agent-workbench 0.5.3 → 0.5.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -3,10 +3,14 @@
3
3
  [![npm](https://img.shields.io/npm/v/agent-workbench)](https://www.npmjs.com/package/agent-workbench)
4
4
  [![CI](https://github.com/cvelasquez/agent-workbench/actions/workflows/ci.yml/badge.svg)](https://github.com/cvelasquez/agent-workbench/actions/workflows/ci.yml)
5
5
  [![license: MIT](https://img.shields.io/npm/l/agent-workbench)](LICENSE)
6
+ [![Mentioned in Awesome Claude Code](https://awesome.re/mentioned-badge.svg)](https://github.com/hesreallyhim/awesome-claude-code)
6
7
 
7
- **One local interface for the Claude Code, Codex, OpenCode and Antigravity
8
- CLIs.** It uses the CLIs you already have logged in, never touches your
9
- credentials, and makes no network calls of its own.
8
+ **English** · [Español](https://github.com/cvelasquez/agent-workbench/blob/main/README.es.md) · [简体中文](https://github.com/cvelasquez/agent-workbench/blob/main/README.zh-CN.md) · [日本語](https://github.com/cvelasquez/agent-workbench/blob/main/README.ja.md) · [Português (Brasil)](https://github.com/cvelasquez/agent-workbench/blob/main/README.pt-BR.md) · [Русский](https://github.com/cvelasquez/agent-workbench/blob/main/README.ru.md) · [한국어](https://github.com/cvelasquez/agent-workbench/blob/main/README.ko.md) · [Français](https://github.com/cvelasquez/agent-workbench/blob/main/README.fr.md) · [Deutsch](https://github.com/cvelasquez/agent-workbench/blob/main/README.de.md)
9
+
10
+ **A visual workbench for the coding agents you already use: the Claude Code,
11
+ Codex, OpenCode and Antigravity CLIs.** It runs on your computer and opens in
12
+ your browser. It uses the CLIs you already have installed and logged in, never
13
+ touches your credentials, and makes no network calls of its own.
10
14
 
11
15
  ```bash
12
16
  npx agent-workbench
@@ -14,89 +18,108 @@ npx agent-workbench
14
18
 
15
19
  ![A tour: the conversation as cards, the changed files and a diff, one tab per CLI, and continuing a conversation with another CLI](https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/demo.webp)
16
20
 
17
- It doesn't talk to any API. It launches the CLI you already have installed and
18
- logged in —`claude`, `codex`, `opencode` or `agy`— inside a pseudo-terminal, and
19
- adds around it what a terminal alone doesn't give you: tabs, browsable history,
20
- the conversation as cards, a context meter, git status and a file tree.
21
+ ## How it works
22
+
23
+ Agent Workbench doesn't talk to any AI service. It starts the CLI you already
24
+ have —`claude`, `codex`, `opencode` or `agy`— inside a real terminal, and adds
25
+ around it what a terminal alone doesn't give you: tabs, your past
26
+ conversations, the conversation as easy-to-read cards, a context meter, the
27
+ files that changed and a file tree.
21
28
 
22
- The terminal is still the terminal. Everything you type reaches the CLI without
23
- the app touching it.
29
+ The terminal is still the terminal: everything you type goes straight to the
30
+ CLI.
24
31
 
25
32
  ---
26
33
 
27
- ## What it does
34
+ ## What you get
28
35
 
29
36
  | | |
30
37
  |---|---|
31
- | **Tabs** | Several live sessions at once, from any of the CLIs, each in its own directory. They survive an `F5`: the processes live on the server, not in the browser tab. Each tab's dot tells you whether the agent is working, idle or waiting for an answer, on the CLIs that publish their status ([below](#multiple-clis)). |
32
- | **Zero-cost startup** | When you open the app, tabs come back as **sleeping tabs**: you can read them in full, and they don't launch any CLI. You open the CLI with a button when you want to write to the agent. |
33
- | **History** | Your projects and past conversations in the sidebar, with a filter, and those of all four CLIs together under each project, each with its badge. Opening one resumes it with its CLI, in the same session. **Archive history…** hides a CLI's sessions from before today in one go, and each project can be archived whole with one button: nothing is deleted, and both can be undone. |
34
- | **Conversation** | The active session's messages, live, with tool calls collapsed and their results inside. A row at the bottom tells you when your message reached the CLI, whether the agent is working and, with Claude Code, which subagents are still at it after the agent finished its turn. Search, jump between matches and back to the end, copy any message or code block and, depending on the CLI, answer the agent's questions from the chat. |
35
- | **Continue with…** | With more than one CLI installed, a conversation can be continued with another CLI in the same folder. The new agent starts from a transcript of the last turns, not from the context the previous one had. |
36
- | **Search everything** | With more than one CLI and something saved in the backup, the sidebar filter also searches the text of every saved conversation, not just their titles. |
37
- | **Context meter** | Tokens from the last request against the model's context window. Tokens, never money. |
38
- | **Changes** | Branch, ahead and behind against the upstream branch, worktrees, and the changed files with their diff. **Read-only.** |
39
- | **Files** | The tree of the tab's directory, with search by name and a preview with syntax highlighting. A button on each row puts its path in your message, without touching the clipboard. A context menu to copy paths, insert them as `@path` or open the file with the system's default app. |
40
- | **Plans** | The documents written by that conversation, rendered: the plans from plan mode, and also the `.md` files the agent created inside the project or in the session's temp folder. Only those named by the conversation you're viewing. |
41
- | **Shared memory** | What agents learn about a project, in `.agents/memory/`, read and written by all four CLIs ([below](#shared-memory)). |
42
- | **Backup** | Optional. The history of all four CLIs in a folder of yours, in the app's own format, so you don't lose it if a CLI changes its format, deletes it, or you uninstall the CLI ([below](#backup-optional)). |
43
- | **Nine languages** | The interface in English, Español, 简体中文, 日本語, Português (Brasil), Русский, 한국어, Français and Deutsch. It follows the browser's language and can be changed in Settings (⚙ in the header), without reloading. |
44
- | **Theme** | Light, dark, or the system's. |
38
+ | **Tabs** | Several agents at once, from any CLI, each in its own folder. A dot on each tab tells you whether the agent is working, idle or waiting for you. Reloading the page doesn't stop anything: the CLIs run on the server, not in the browser. |
39
+ | **Fast start** | When you open the app, your tabs come back **asleep**: you can read them in full, and no CLI starts until you click **Open CLI**. |
40
+ | **History** | Your projects and past conversations in the sidebar, from every CLI, with a filter. Click one to pick it up where you left it, with its CLI. Archive old sessions or a whole project to tidy up: nothing is deleted, and it can be undone. |
41
+ | **Conversation** | The messages as cards, live, with tool calls folded. Older messages load as you scroll up. Search, copy any message or code block, and jump back to your previous message (↑) or to the end (↓). A row at the bottom says when your message reached the CLI and whether the agent is still working. Depending on the CLI, you can answer the agent's questions right there. |
42
+ | **Message box** | Room to write, paste images, attach files, and long pasted text kept as a tidy chip. What you're typing survives closing the app, and shows up live in every open window. |
43
+ | **Context meter** | How much of the model's context window the conversation is using. Tokens, never money. |
44
+ | **Changes** | The branch, the changed files and their diff. **Read-only**: no commit or push buttons. |
45
+ | **Files** | The project's file tree, with search and a preview with syntax highlighting. Open several files at once, each in its own tab; Markdown files open formatted. One click puts a file's path in your message. |
46
+ | **Plans** | The plans and `.md` documents the conversation wrote, rendered. |
47
+ | **Notes** | A scratchpad next to the conversation, images included. With Claude Code, one click sends a note to a new conversation. |
48
+ | **Console** | Your system's terminal at the bottom, in the project's folder: PowerShell, cmd or Git Bash on Windows; bash or zsh on macOS and Linux. |
49
+ | **Continue with…** | With more than one CLI, continue a conversation with another one in the same folder. The new agent starts from a transcript of the last turns. |
50
+ | **Notifications** | A sound, and optional system notifications, when an agent finishes or is waiting for you. |
51
+ | **Settings** | ⚙ in the header: the language, the terminal's theme, which panels show, which console opens and which CLIs the app uses. |
52
+ | **Nine languages** | English, Español, 简体中文, 日本語, Português (Brasil), Русский, 한국어, Français and Deutsch. It follows your browser's language. |
53
+ | **Theme** | Light, dark or the system's. The terminal stays dark by default, so the CLIs' colors stay readable. |
54
+
55
+ Optional, and off by default: [shared memory](#shared-memory-between-clis)
56
+ between CLIs, a [backup](#backup-of-your-history) of your history,
57
+ [remote access](#remote-access-over-ssh) from another computer or a phone, and
58
+ an [Android app](#the-android-app).
45
59
 
46
60
  <p>
47
61
  <img src="https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/captura-cambios.png" width="49%" alt="The Changes panel: branch and changed files, by group">
48
62
  <img src="https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/captura-diff.png" width="49%" alt="The diff of one of those files, in the same panel">
49
63
  </p>
50
64
 
51
- ### Multiple CLIs
65
+ ### With more than one CLI
52
66
 
53
- With more than one CLI installed, every session in the sidebar and every tab
54
- carries its CLI's badge, and the new-tab `+` opens with the one you used in that
55
- project; its arrow lets you pick another. With just one, you see none of this.
67
+ Each session and each tab shows a small badge with its CLI, and the `+` for a
68
+ new tab lets you pick which one to open. With a single CLI you see none of
69
+ this.
56
70
 
57
71
  <img src="https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/captura-clis.png" width="45%" alt="The projects sidebar with sessions from several CLIs, each with its badge, and the + menu with the four CLIs and their versions">
58
72
 
59
- **Not every CLI exposes the same things.** Claude Code leaves everything the app
60
- needs in its files. Codex leaves the history and the meter's tokens, but not its
61
- status, its pending permissions, its questions or the model: its dot says the
62
- status is unknown, there's no "waiting" notice and questions are answered in its
63
- terminal. With OpenCode, the app starts a local OpenCode server when you open a
64
- tab, and that's where its status, the notice that it's waiting for a permission
65
- and the questions you answer from the chat come from. Antigravity CLI publishes
66
- its status and tokens only if you configure its status line
67
- ([below](#antigravity-cli-status-and-meter-optional)).
73
+ **Not every CLI shares the same information**, so a few features depend on the
74
+ CLI:
75
+
76
+ | | Claude Code | Codex | OpenCode | Antigravity CLI |
77
+ |---|:---:|:---:|:---:|:---:|
78
+ | Tab dot: working, idle, waiting for you | ✓ | — | ✓ ¹ | ✓ ² |
79
+ | Answer the agent's questions from the conversation | ✓ | — | ✓ ¹ | — |
80
+ | Context meter | ✓ | ✓ | ✓ | ✓ ² |
81
+ | Mode and model pickers | ✓ | — | — | ✓ |
82
+ | Paste images | ✓ | ✓ | ✓ | — |
83
+ | Plans tab | ✓ | — | — | — |
84
+ | Send a note to a new conversation | ✓ | — | — | — |
85
+
86
+ The CLI's own terminal is always there for the rest: where there's a "—",
87
+ questions, for example, are answered in it.
88
+
89
+ ¹ With OpenCode, the app starts a local OpenCode server to get this
90
+ ([details](#your-data-and-security)).
91
+ ² Only if you set up its status line
92
+ ([how](#antigravity-cli-status-and-context-meter)).
68
93
 
69
94
  ---
70
95
 
71
96
  ## Requirements
72
97
 
73
- | | |
74
- |---|---|
75
- | **Node.js** | 20 or later. **To see the OpenCode history and the Antigravity CLI titles, 22.13 or later**: they're read with the SQLite that Node ships since that version. With an older one everything else works the same; the startup output warns about OpenCode, and Antigravity CLI lists its conversations without the titles or folders from its index |
76
- | **git** | for the Changes panel; everything else works without it |
77
- | **At least one CLI** | installed and logged in (table below) |
98
+ - **Node.js 20 or later.** To see OpenCode's history and Antigravity CLI's
99
+ conversation titles, **22.13 or later**: they're read with the SQLite built
100
+ into Node since that version. With an older one, everything else works.
101
+ - **git**, for the Changes panel. Everything else works without it.
102
+ - **At least one CLI**, installed and logged in:
78
103
 
79
- | CLI | Command | Installation |
104
+ | CLI | Command | How to install |
80
105
  |---|---|---|
81
106
  | Claude Code | `claude` | [installation guide](https://docs.claude.com/en/docs/claude-code/setup) |
82
107
  | Codex | `codex` | [guide](https://learn.chatgpt.com/docs/codex/cli) |
83
108
  | OpenCode | `opencode` | [documentation](https://opencode.ai/docs/) |
84
- | Antigravity CLI | `agy` | [guide](https://antigravity.google/docs/cli/getting-started). For its status and meter, also `node` in the CLI's `PATH` |
109
+ | Antigravity CLI | `agy` | [guide](https://antigravity.google/docs/cli/getting-started) |
85
110
 
86
- Agent Workbench does **not** bundle or download any CLI: it uses the ones you
87
- already have in your `PATH`. If it doesn't find any, it tells you so and doesn't
88
- open sessions.
111
+ Agent Workbench doesn't include or download any CLI: it finds the ones in your
112
+ `PATH`. If there are none, it tells you so and doesn't open sessions.
89
113
 
90
- Developed and used daily on Windows 11 with PowerShell, which is the main
91
- platform. On macOS and Linux, CI installs the package and starts it on every
92
- push —server up, a pseudo-terminal spawning a process—, but it sees less daily
93
- use there: reports are welcome. On Linux, the `node-pty` dependency doesn't ship a prebuilt
94
- binary and is compiled on install: you need `python3`, `make` and a C++ compiler
95
- (`build-essential` on Debian and Ubuntu).
114
+ **Platforms.** Built and used every day on Windows 11 with PowerShell. On macOS
115
+ and Linux, every change is installed and started automatically, but they get
116
+ less daily use: reports are welcome. On Linux, installing compiles the terminal
117
+ module, so you need `python3`, `make` and a C++ compiler (`build-essential` on
118
+ Debian and Ubuntu).
96
119
 
97
120
  ---
98
121
 
99
- ## Install
122
+ ## Install and start
100
123
 
101
124
  ```bash
102
125
  npm install -g agent-workbench
@@ -108,110 +131,143 @@ Then, from the folder of the project you want to work on:
108
131
  agent-workbench
109
132
  ```
110
133
 
111
- To try it once without installing it, `npx agent-workbench` — it downloads about
112
- 60 MB whenever the npm cache is cold, mostly the terminal's native binary. For
113
- daily use, the global install is the better choice.
114
-
115
- The server prints a URL with a token and opens it in the browser:
134
+ It opens your browser at an address like this one:
116
135
 
117
136
  ```
118
137
  URL http://127.0.0.1:52341/?token=…
119
138
  ```
120
139
 
121
- That URL is the only way in. The token is different on every start, and the
122
- server listens only on `127.0.0.1`.
140
+ That address is the only way in: the token changes on every start, and the app
141
+ only listens on your own computer (`127.0.0.1`).
142
+
143
+ To try it once without installing: `npx agent-workbench`. It downloads about
144
+ 60 MB the first time, mostly the terminal's native module. For everyday use,
145
+ the global install is better.
146
+
147
+ ---
148
+
149
+ ## Keyboard shortcuts
150
+
151
+ | Keys | Action |
152
+ |---|---|
153
+ | `Alt+T` | New tab |
154
+ | `Alt+W` | Close the tab |
155
+ | `Alt+←` / `Alt+→` (or `Alt+PgUp` / `Alt+PgDn`) | Previous / next tab |
156
+ | `Alt+P` | Show or hide the right panel |
157
+ | `Shift+Tab`, outside the terminal | Back to the previous tab |
158
+
159
+ The `?` button in the header lists them all, along with the active CLI's own
160
+ shortcuts. They use `Alt` because the browser keeps `Ctrl+T`, `Ctrl+W` and
161
+ `Ctrl+Tab` for its own tabs and never passes them to the page.
162
+
163
+ Every other key —`Esc`, `Ctrl+C`, `Ctrl+R`, `Ctrl+O`, the arrows, `Alt+V` to
164
+ paste images— reaches the CLI untouched. `Ctrl+V` too: what it does there is up
165
+ to the CLI. Only in the console at the bottom, on Windows, `Ctrl+V` pastes text,
166
+ as in Windows Terminal.
167
+
168
+ ---
169
+
170
+ ## Optional extras
123
171
 
124
- ### Shared memory
172
+ ### Shared memory between CLIs
125
173
 
126
- Each CLI keeps what it learns about a project in its own folder, and the others
127
- can't see it. The **Memory** tab installs a bridge so all four use the same one:
128
- notes in the project's `.agents/memory/`, which each CLI reads and writes —from
129
- here or from its own terminal— through `AGENTS.md` and `CLAUDE.md`.
174
+ Each CLI keeps what it learns about a project in its own place, and the others
175
+ can't see it. The **Memory** tab sets up one memory for all four: notes in the
176
+ project's `.agents/memory/` folder, which every CLI reads and writes —from here
177
+ or from its own terminal— through `AGENTS.md` and `CLAUDE.md`. The tab is hidden
178
+ at first: show it from ⚙ Settings.
130
179
 
131
- **Preview changes** shows, file by file, what will be written, and nothing is
132
- touched until you confirm. Installing imports the memory Claude Code already had
133
- for that project. The app doesn't write each CLI's global memory: it gives you
134
- the snippet to paste yourself.
180
+ **Preview changes** shows, file by file, what would be written, and nothing is
181
+ touched until you confirm. Setting it up imports the memory Claude Code already
182
+ had for that project. The app doesn't edit each CLI's global settings: it gives
183
+ you the snippet to paste yourself.
135
184
 
136
185
  <img src="https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/captura-memoria.png" width="70%" alt="The Memory tab with the bridge installed for the four CLIs and the imported notes">
137
186
 
138
- ### Antigravity CLI: status and meter (optional)
187
+ ### Backup of your history
139
188
 
140
- Antigravity CLI doesn't record in any file whether it's working, waiting for you
141
- to authorize a tool or idle, nor how many tokens it has used: it only publishes
142
- that through its *status line*. Without it configured, its tabs work the same
143
- —history, conversation, mode, model— but the tab's dot says the status is
144
- unknown and the meter stays empty. To turn it on:
189
+ Each CLI owns its history, and it can change its format, prune it, or be
190
+ uninstalled. The backup keeps a copy of your conversations —messages, tool
191
+ calls and results, images, and each project's memory— in a folder of yours, in
192
+ files you can read without the app.
145
193
 
146
- 1. Open an Antigravity tab and click **Configure**, next to the meter.
147
- 2. Copy the line the dialog shows and merge it into whatever
148
- `~/.gemini/antigravity-cli/settings.json` already has. The app doesn't touch
149
- that file: you edit it.
150
- 3. The dialog switches to **Configured** on its own within a couple of seconds,
151
- or with **Check**.
152
-
153
- The line runs a script the app keeps in its own folder. It saves only each
154
- conversation's status, mode, model and tokens, in that same folder; it doesn't
155
- save your email, quota, plan or cost, which the CLI also passes to it, and it
156
- prints nothing, so the CLI's own status line stays as it is. Once set, **every**
157
- `agy` session runs it, including those you open outside the app, and it needs
158
- `node` in the `PATH`. On Windows the line changes into the script's folder
159
- instead of naming the script in quotes: the CLI runs it with `cmd /c`, and no
160
- quote reaches `node` intact.
161
-
162
- ### Backup (optional)
163
-
164
- Each conversation's history belongs to its CLI, in its format, and a CLI can
165
- change it, prune it or cease to exist. The backup keeps the same as that
166
- history —messages, tool inputs and results, images and each project's memory—
167
- in a folder of yours, in files you can read without the app. **It's off by
168
- default.** The backup button, in the header of the projects sidebar, opens a
169
- dialog: first **Measure** tells you how much space it would take per CLI,
170
- without writing anything, and then **Turn on** enables it. Once on, it copies
171
- everything the sidebar lists that isn't archived, and copies each session that
172
- changes again a minute after it goes quiet. Archived sessions aren't copied, and
173
- archiving doesn't delete what was already copied: the app never deletes anything
194
+ **It's off by default.** Open it with the backup button at the top of the
195
+ projects sidebar, or from ⚙ Settings. **Measure** tells you how much space it
196
+ would take per CLI, without writing anything, and **Turn on** starts it. From
197
+ then on it copies everything in the sidebar that isn't archived, and copies a
198
+ conversation again a minute after it stops changing. It never deletes anything
174
199
  from that folder.
175
200
 
176
- Whatever the CLI no longer has stays in the sidebar, marked as a backup, and
177
- opens in Markdown; each project can be exported to Markdown. And with the backup
178
- on and more than one CLI, the sidebar filter offers **In conversations**: it
179
- searches the text of everything copied, from every CLI.
201
+ With the backup on:
202
+
203
+ - Conversations a CLI no longer has stay in the sidebar, marked as backup, and
204
+ open as Markdown.
205
+ - Any project can be exported to Markdown.
206
+ - With more than one CLI, the sidebar filter can search **In conversations**:
207
+ the text of everything copied, from every CLI.
180
208
 
181
209
  <img src="https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/captura-buscador.png" width="40%" alt="Search in conversations: a match in one session of each CLI, with its snippet">
182
210
 
183
- By default it lives inside the app's configuration folder
184
- (`%APPDATA%\agent-workbench\vault` on Windows). **Change folder…** copies all of
185
- it to another one —a synced folder, another drive— without overwriting anything,
186
- and the previous one stays as it was. If you put it in a synced folder or in a
187
- repository, what it keeps travels with it.
211
+ By default it lives in the app's configuration folder
212
+ (`%APPDATA%\agent-workbench\vault` on Windows). **Change folder…** copies it to
213
+ another one —a synced folder, another drive— without overwriting anything, and
214
+ leaves the old one as it was. If you put it in a synced folder or a repository,
215
+ your conversations go with it.
216
+
217
+ <details>
218
+ <summary>Importing history from Gemini CLI and the Antigravity IDE</summary>
188
219
 
189
- Two one-off importers bring in history from tools the app doesn't read. They run
190
- from source ([below](#from-source)) and **write nothing without `--write`**:
191
- without it, they tell you what they would import.
220
+ Two one-time importers bring in history from tools the app doesn't read. They
221
+ run from the source code ([below](#from-source)) and **write nothing without
222
+ `--write`**: without it, they only tell you what they would import.
192
223
 
193
224
  - `pnpm vault:import gemini-cli [--cwd <folder>]` — the chats left over from
194
- Gemini CLI; `--cwd` names the folder where you used it, to place them in their
225
+ Gemini CLI. `--cwd` is the folder where you used it, to file them under that
195
226
  project.
196
- - `pnpm vault:import antigravity-ide --workspace <folder>` — whatever is
197
- readable from the Antigravity IDE conversations in that folder: each one's
198
- summary and its `.md` documents. The conversation content is encrypted, so it
199
- is marked as partial history.
227
+ - `pnpm vault:import antigravity-ide --workspace <folder>` — what can be read
228
+ from the Antigravity IDE conversations in that folder: each one's summary and
229
+ its `.md` documents. The conversation itself is encrypted, so it's marked as
230
+ partial history.
200
231
 
201
- ### Remote access over SSH (optional)
232
+ </details>
202
233
 
203
- You can leave Agent Workbench running on one computer and use it from another
204
- one on the same network, in its browser — with no CLI installed there.
205
- **The app is never opened to the network**: it keeps listening only on
206
- `127.0.0.1`, and the other computer reaches it through an SSH tunnel. That's
207
- also what makes it work: through the tunnel the other browser sees the app as
208
- `localhost`, and outside `localhost` browsers switch off things the app needs
209
- (the clipboard, notifications, `crypto.randomUUID`). **It's off by default.**
234
+ ### Antigravity CLI: status and context meter
210
235
 
211
- **1. On the computer that runs the app, once: turn on its SSH server.** The app
212
- doesn't do this for you — it's a system change that needs administrator rights.
236
+ Antigravity CLI only tells whether it's working or waiting for you, and how
237
+ many tokens it has used, through its *status line*. Without it, its tabs work
238
+ normally, but the tab's dot says the status is unknown and the meter stays
239
+ empty. To turn it on:
213
240
 
214
- - **Windows 10/11** — in PowerShell *as administrator*:
241
+ 1. Open an Antigravity tab and click **Configure**, next to the meter.
242
+ 2. Copy the line it shows and add it to what
243
+ `~/.gemini/antigravity-cli/settings.json` already has. The app doesn't edit
244
+ that file: you do.
245
+ 3. Within a couple of seconds, or with **Check**, the dialog says
246
+ **Configured**.
247
+
248
+ The line runs a small script kept in the app's own folder. It saves only each
249
+ conversation's status, mode, model and tokens —not your email, quota, plan or
250
+ cost, which the CLI also passes to it— and prints nothing, so the CLI's own
251
+ status line stays as it was. Once set, it runs on **every** `agy` session,
252
+ including the ones you open outside the app, and it needs `node` in the
253
+ `PATH`.
254
+
255
+ ### Remote access over SSH
256
+
257
+ Leave Agent Workbench running on one computer and use it from another one on
258
+ your network, in its browser, with no CLI installed there. **The app is never
259
+ opened to the network**: it still listens only on `127.0.0.1`, and the other
260
+ computer reaches it through an SSH tunnel. The tunnel is also what makes it
261
+ work: through it the other browser sees the app as `localhost`, which browsers
262
+ require for the clipboard and notifications. **It's off by default.**
263
+
264
+ **1. On the computer that runs the app, once: turn on its SSH server.** It's a
265
+ system change that needs administrator rights, so the app doesn't do it for
266
+ you.
267
+
268
+ - **Windows 10/11** — in PowerShell *as administrator*. The first line downloads
269
+ and installs the server and takes a few minutes: wait until the prompt comes
270
+ back.
215
271
 
216
272
  ```powershell
217
273
  Add-WindowsCapability -Online -Name OpenSSH.Server~~~~0.0.1.0
@@ -219,93 +275,87 @@ doesn't do this for you — it's a system change that needs administrator rights
219
275
  Set-Service -Name sshd -StartupType Automatic
220
276
  ```
221
277
 
222
- The first line downloads and installs the server: it takes a few minutes and
223
- shows "Running" meanwhile. Wait until the prompt comes back.
224
-
225
278
  - **macOS** — System Settings → General → Sharing → **Remote Login**.
226
- - **Linux** — install and enable your distribution's OpenSSH server
227
- (`sudo apt install openssh-server`, then `sudo systemctl enable --now ssh`).
279
+ - **Linux** — install and turn on your distribution's OpenSSH server. On Debian
280
+ or Ubuntu:
228
281
 
229
- **2. In the app: the ⇄ button in the header → Turn on remote access**, and
230
- restart Agent Workbench. A tunnel needs a port that doesn't change on every
231
- start, so the app uses a fixed one (24837 by default, you can change it) — and
232
- the port is chosen at startup, which is why it takes a restart. Restarting
233
- closes the open CLIs; the tabs come back asleep. Turning it *off* takes effect
234
- right away.
282
+ ```bash
283
+ sudo apt install openssh-server
284
+ sudo systemctl enable --now ssh
285
+ ```
286
+
287
+ **2. In the app: ⇄ in the header → Turn on remote access, then restart Agent
288
+ Workbench.** The tunnel needs a port that doesn't change, so the app uses a
289
+ fixed one (24837 by default; you can change it), and the port is chosen at
290
+ startup. Restarting closes the open CLIs; your tabs come back asleep. Turning
291
+ it *off* works right away.
235
292
 
236
- **3. On the other computer, every time: open the tunnel.** The dialog shows the
293
+ **3. On the other computer, each time: open the tunnel.** The dialog shows the
237
294
  command ready to copy:
238
295
 
239
296
  ```
240
297
  ssh -N -L 24837:127.0.0.1:24837 you@your-computer
241
298
  ```
242
299
 
243
- Leave it running. **Use the same port on both ends**: the server checks it, and
244
- with a different local port it answers 403. Adding
300
+ Leave it running, and **use the same port on both ends**: with a different
301
+ local port the app answers 403. Adding
245
302
  `-o ServerAliveInterval=30 -o ExitOnForwardFailure=yes` makes the tunnel notice
246
- a dropped Wi‑Fi instead of hanging. On Windows you can save the line in a
303
+ a dropped Wi‑Fi instead of hanging. On Windows you can save the command in a
247
304
  `.cmd` file and double-click it.
248
305
 
249
- **4. Pair that computer, once.** In the dialog, **Pair a device** shows an
250
- address with a one-time code, such as
251
- `http://localhost:24837/?pair=K7QM-X2RD`. Open it in the other computer's
252
- browser within five minutes. From then on that browser just opens
253
- `http://localhost:24837`, even after the app restarts. The dialog lists the
254
- paired devices, and **Revoke** cuts one off on the spot.
255
-
256
- A paired device can use the whole app, with two limits: it can't manage remote
257
- access (turn it on, pair or revoke), and it isn't offered what opens a window on
258
- the computer that runs the app — *Open with the default app*, opening the local
259
- copy's folder or exporting to Markdown. A small **Remote** badge in the header
260
- says so. Those limits are about what lands on the host's screen, not about
261
- trust: the app includes a console and launches agents, so a paired device can
262
- run commands on that computer, just as you can. Pair only devices that are
263
- yours.
264
-
265
- **A safer key (recommended).** The SSH login is the one of your user on that
266
- computer, so a password or key that opens the tunnel also opens a full terminal
267
- there. You can create a key that is only good for the tunnel: generate one on
268
- the other computer (`ssh-keygen -t ed25519`) and add its public half to
269
- `authorized_keys` on the computer that runs the app, with these options in
270
- front:
306
+ **4. Pair that computer, once.** In the dialog, **Pair a device** gives you an
307
+ address with a one-time code, like `http://localhost:24837/?pair=K7QM-X2RD`.
308
+ Open it in the other computer's browser within five minutes. From then on, that
309
+ browser just opens `http://localhost:24837`, even after the app restarts. The
310
+ dialog lists the paired devices, and **Revoke** cuts one off at once.
311
+
312
+ A paired device can use the whole app, except managing remote access and the
313
+ few things that would open a window on the host computer, like *Open with the
314
+ default app*. A **Remote** badge in the header shows it. Keep in mind that the
315
+ app includes a console and launches agents, so a paired device can run commands
316
+ on your computer, just as you can. **Pair only devices that are yours.**
317
+
318
+ **On a phone or tablet.** Through the same tunnel, a phone's browser gets a
319
+ one-column layout: the projects sidebar slides in from the side, the tabs sit
320
+ in the header, and a strip switches between the conversation, the CLI, changes,
321
+ files and the rest. It appears on any window narrower than 768 px. On a touch
322
+ screen, Enter in the message box starts a new line (the send button sends), and
323
+ dragging a finger over a terminal scrolls it.
324
+
325
+ <details>
326
+ <summary>A safer SSH key, only good for the tunnel (recommended)</summary>
327
+
328
+ The SSH login is your user's on that computer, so a password or key that opens
329
+ the tunnel also opens a full terminal there. You can create a key that is only
330
+ good for the tunnel: generate one on the other computer
331
+ (`ssh-keygen -t ed25519`) and add its public half to `authorized_keys` on the
332
+ computer that runs the app, with these options in front:
271
333
 
272
334
  ```
273
335
  restrict,port-forwarding,permitopen="127.0.0.1:24837",command="exit" ssh-ed25519 AAAA… other-computer
274
336
  ```
275
337
 
276
- `restrict` on its own still lets the key run commands (it only refuses a
277
- terminal, agent forwarding and other ports); `command="exit"` is what closes
278
- that door, and the tunnel never asks for a command. On Windows, if your user is
279
- an administrator, that file is
280
- `C:\ProgramData\ssh\administrators_authorized_keys` and it must be writable
338
+ `restrict` alone still lets the key run commands; `command="exit"` is what
339
+ closes that door, and the tunnel never asks for a command. On Windows, if your
340
+ user is an administrator, that file is
341
+ `C:\ProgramData\ssh\administrators_authorized_keys`, and it must be writable
281
342
  only by Administrators and SYSTEM; otherwise it's
282
- `%USERPROFILE%\.ssh\authorized_keys`. To check that the restriction holds, try
343
+ `%USERPROFILE%\.ssh\authorized_keys`. To check it, try
283
344
  `ssh you@your-computer whoami` with that key: it should print nothing and end
284
345
  at once, while the `-N -L` tunnel keeps working.
285
346
 
286
- **From a phone.** Through the same tunnel, a phone browser gets a
287
- single-column layout: the projects sidebar as a drawer, the tabs in the header,
288
- and a strip that switches between the conversation, the CLI, changes, files and
289
- the rest. It appears on any window narrower than 768 px (a tablet held upright, too).
290
- On a touch screen, Enter in the message box starts a new line (the send button
291
- sends), and dragging a finger over a terminal scrolls it like a mouse wheel.
292
-
293
- **Notifications.** The bell button in the header turns on system notifications:
294
- "finished" and "is waiting for you", with the tab's name, only when the window
295
- isn't in view. They follow the same rule as the notification sound, so they
296
- come from the CLIs that publish their status. Each browser remembers its own
297
- choice, and the sound plays on every screen that has the app open.
347
+ </details>
298
348
 
299
- ### The Android app (optional)
349
+ ### The Android app
300
350
 
301
351
  An app for Android 10 and later opens the tunnel for you and shows the same
302
- interface, served by your computer, in the phone-sized layout. It keeps the
303
- connection open in the background and notifies you when an agent finishes or is
304
- waiting for you, with the screen off too; tapping a notification opens that
305
- tab. When the Wi‑Fi comes back, it reconnects on its own.
352
+ interface, served by your computer, in the phone layout. It stays connected in
353
+ the background and notifies you when an agent finishes or is waiting for you,
354
+ even with the screen off; tapping a notification opens that tab. When the Wi‑Fi
355
+ comes back, it reconnects on its own.
306
356
 
307
357
  It needs remote access on and the computer's SSH server running (steps 1 and 2
308
- above). Install it from
358
+ above). Get it on
309
359
  [Google Play](https://play.google.com/store/apps/details?id=io.github.cvelasquez.agentworkbench),
310
360
  or build it yourself from [`android/`](android/) as
311
361
  [`CONTRIBUTING.md`](CONTRIBUTING.md#the-android-app) explains.
@@ -315,18 +365,17 @@ or build it yourself from [`android/`](android/) as
315
365
  1. In the ⇄ dialog, choose **Pair a phone**. It shows a line to paste in a
316
366
  terminal on this computer, and a QR code. On Windows, if your user is an
317
367
  administrator, the line goes in PowerShell *as administrator*; the dialog
318
- says so.
319
- 2. Paste the line. It adds the phone's key to `authorized_keys`, already
320
- restricted to the tunnel: with that key the phone reaches the app's port and
321
- nothing else — no terminal and no commands.
322
- 3. Scan the QR with the app. The code works once and for five minutes, and the
323
- dialog has to stay open until the phone connects.
324
-
325
- The QR carries the phone's private SSH key. This computer generates it, keeps
326
- it only in memory until the phone uses it and never writes it to disk: don't
327
- share a picture of that QR. The phone saves your computer's SSH fingerprint on
328
- the first connection, and if it ever changes, the app refuses to connect and
329
- tells you.
368
+ tells you.
369
+ 2. Paste the line. It authorizes the phone's key for the tunnel only: with it,
370
+ the phone reaches the app and nothing else — no terminal, no commands.
371
+ 3. Scan the QR code with the app. The code works once and for five minutes, and
372
+ the dialog has to stay open until the phone connects.
373
+
374
+ The QR code carries the phone's private SSH key. Your computer keeps it only in
375
+ memory until the phone uses it and never writes it to disk: **don't share a
376
+ picture of that QR code**. The phone remembers your computer's SSH fingerprint
377
+ on the first connection, and if it ever changes, it refuses to connect and tells
378
+ you.
330
379
 
331
380
  **To remove a phone**, **Revoke** it in the dialog and delete its line from
332
381
  `authorized_keys` (it ends in `agent-workbench-phone`). On the phone,
@@ -339,132 +388,83 @@ app keeps on the phone and the permissions it asks for are in
339
388
 
340
389
  ---
341
390
 
342
- ## From source
343
-
344
- To work on the app, or if you'd rather not install anything globally:
345
-
346
- ```bash
347
- corepack enable pnpm
348
- pnpm install
349
- pnpm dev # Vite with hot reload
350
- ```
351
-
352
- ```bash
353
- pnpm build # builds the interface once
354
- pnpm start # serves the built interface, without Vite
355
- ```
356
-
357
- ### One-click start on Windows
358
-
359
- ```bash
360
- pnpm package
361
- ```
362
-
363
- It builds the interface and leaves an **`Agent Workbench.cmd`** in the root.
364
- Double-click it and you're done: it installs whatever is missing, builds if
365
- needed and opens the browser. It's a twenty-line text file; you can read all of
366
- it before running it.
367
-
368
- ---
369
-
370
- ## Keyboard shortcuts
371
-
372
- `Alt+T` new tab · `Alt+W` close · `Alt+←/→` (or `Alt+PgUp/PgDn`) switch tabs ·
373
- `Alt+P` show or hide the right panel · `Shift+Tab`, outside the terminal, go
374
- back to the previous tab.
375
-
376
- The `?` button in the top bar lists them all, along with the shortcuts of the
377
- tab's CLI.
378
-
379
- **Why `Alt` and not `Ctrl`:** the browser keeps `Ctrl+T`, `Ctrl+W` and
380
- `Ctrl+Tab` for its own tabs and the event never reaches the page. It's not
381
- something `preventDefault` can fix: there's no event to prevent.
382
-
383
- The app captures exactly those combinations and no others. `Shift+Tab` only
384
- when focus isn't on the terminal, because inside it's the key the CLI uses to
385
- switch modes. Everything else —`Esc`, `Esc Esc`, `Ctrl+C`, `Ctrl+R`, `Ctrl+O`,
386
- the arrow keys and above all `Alt+V`, which pastes images— reaches the CLI
387
- untouched. `Ctrl+V` too: what it does there is up to the CLI. Only in the
388
- console at the bottom, on Windows, `Ctrl+V` pastes text, as in Windows
389
- Terminal.
390
-
391
- ---
392
-
393
- ## What the app does with your data
394
-
395
- Nothing leaves your machine. No telemetry, no analytics, no outgoing network
396
- calls at all. The exception is the CLI itself at work: the one you open in a
397
- tab, and the OpenCode server described below, talk to the model provider just
398
- as they would if you opened them by hand. The Android app talks only to your
399
- computer; [`PRIVACY.md`](PRIVACY.md) lists what it keeps on the phone.
400
-
401
- **It never touches your credentials.** There's no login in the interface: if
402
- you aren't logged in, you log in inside the CLI's terminal and the app doesn't
403
- even notice.
391
+ ## Your data and security
392
+
393
+ - **Nothing leaves your computer.** No telemetry, no analytics, no network calls
394
+ of its own. Only the CLIs talk to their AI provider, just as they would if you
395
+ opened them by hand. The Android app talks only to your computer.
396
+ - **It never touches your credentials.** There's no login screen: if a CLI asks
397
+ you to log in, you do it in its own terminal, and the app doesn't even notice.
398
+ - **It reads the CLIs' history and writes nothing in their folders.** The table
399
+ below lists what it reads and what it never opens.
400
+ - **It writes only in its own folders**, plus one thing in your projects, and
401
+ only if you confirm it: the shared memory.
402
+ - **The server listens only on your computer** (`127.0.0.1`), on a random port,
403
+ with a new random token on every start, and it rejects requests coming from
404
+ other web pages. Remote access doesn't change that: the port becomes a fixed
405
+ one, still on `127.0.0.1`, and SSH —which you turn on and control— does the
406
+ rest.
407
+ - **The Changes panel is read-only.** No commit, stage or push: with an agent
408
+ editing files, a button that writes history is exactly the kind of thing
409
+ where, later, nobody knows who pressed it.
410
+
411
+ | CLI | What it reads | What it never opens |
412
+ |---|---|---|
413
+ | Claude Code | From `~/.claude/`: `projects/` (the history, and each project's memory, to import it), `sessions/` (whether the CLI is waiting for you) and `plans/`. For the Plans tab, also the `.md` files the conversation wrote in the project or in that session's temp folder — **only the ones named by the conversation you're viewing**, without browsing any folder | `.credentials.json` or any token |
414
+ | Codex | From `~/.codex/` (or `CODEX_HOME`): `sessions/` and `archived_sessions/` | `auth.json`, `config.toml` or its `*.sqlite` databases |
415
+ | OpenCode | Its `opencode.db` database, opened read-only, and from it only the sessions, messages and parts; and its model catalog, for the context window size | `auth.json`, `opencode.json`, or the accounts, credentials, permissions and shared-sessions tables |
416
+ | Antigravity CLI | From `~/.gemini/antigravity-cli/`: each conversation's transcripts, `history.jsonl`, the last conversation of each folder, the model and status line from `settings.json`, and a temporary **copy** of its conversation index; from `~/.gemini/config/projects/`, each project's folder | Its MCP configuration, its entry in the system keychain, the contents of `conversations/`, and `~/.gemini/antigravity/`, which is its IDE |
404
417
 
405
- **From each CLI it reads only this, and it writes nothing in the CLI's
406
- folder.** The backup importers, when you run them, also read what
407
- [their section](#backup-optional) says.
418
+ The [backup importers](#backup-of-your-history), when you run them, also read
419
+ what their section says. The full list, file by file, is in
420
+ [`ARCHITECTURE.md`](ARCHITECTURE.md#4-what-is-read-from-each-cli).
408
421
 
409
- | CLI | Reads | Never opens |
410
- |---|---|---|
411
- | Claude Code | from `~/.claude/`: `projects/` (the history, and each project's memory, to import it), `sessions/` (whether the CLI is waiting for an answer) and `plans/`. And for the Plans tab, the `.md` files the conversation wrote inside the project or in that session's temp folder: **only those named by the conversation you're viewing**, without walking any folder | `.credentials.json` or any token |
412
- | Codex | from `~/.codex/` (or `CODEX_HOME`): `sessions/` and `archived_sessions/` | `auth.json`, `config.toml` or its `*.sqlite` databases |
413
- | OpenCode | its `opencode.db` database, opened read-only, and from it only the sessions, messages and parts tables; and its model catalog, for the context window size | `auth.json`, `opencode.json`, or the accounts, credentials, permissions and shared-sessions tables |
414
- | Antigravity CLI | from `~/.gemini/antigravity-cli/`: each conversation's transcripts, `history.jsonl`, the last conversation of each folder, from `settings.json` only the model and the status line, and its conversation index, from a temporary **copy**; from `~/.gemini/config/projects/`, each project's folder | the MCP configuration, its entry in the system keychain, the contents of `conversations/`, or `~/.gemini/antigravity/`, which is its IDE |
422
+ <details>
423
+ <summary>The fine print: footprints, environment variables and what the app writes</summary>
415
424
 
416
- Three footprints worth knowing about:
425
+ **Three footprints worth knowing about:**
417
426
 
418
427
  - **Reading the OpenCode database** does what SQLite does with any reader: it
419
- creates its `-wal` and `-shm` files if they're missing and updates the
420
- timestamp of `-shm`. It never runs an OpenCode command to read it.
421
- - **If an Antigravity tab's own log doesn't show up**, it reads the CLI's
428
+ creates its `-wal` and `-shm` files if they're missing and updates the date of
429
+ `-shm`. The app never runs an OpenCode command to read it.
430
+ - **If an Antigravity tab's own log doesn't show up**, the app reads the CLI's
422
431
  `log/cli-*.log` files, which contain your messages and the account's email,
423
- only to find the conversation id, and without keeping any line.
424
- - **With OpenCode, the app runs its local server**, `opencode serve`: just one,
425
- from when you open the first OpenCode tab until five minutes after you close
426
- the last one, or until you close the app. It listens on `127.0.0.1`, with an
427
- ephemeral port and a different password on every start, even if your OpenCode
428
- configuration says otherwise. The app only asks it for the sessions' status,
429
- pending permissions and questions, to create a session, to answer a question
430
- and to abort a session: nothing about your configuration or your accounts.
431
-
432
- **It adds no authentication variables** to the environment of the CLIs it
433
- launches. The environment is inherited as is, with two exceptions: it
434
- **removes** `CLAUDE_CODE_CHILD_SESSION` —which turns off history saving— and
435
- tells you with a banner when it does; and it **adds** a single variable to the
436
- OpenCode server, `OPENCODE_SERVER_PASSWORD`, holding that per-start password. It
437
- isn't any account's password, and the variable doesn't reach any tab: each
438
- OpenCode tab gets the password on its command line (`attach --password`), where
439
- other processes running as your user can see it. It only works for that server
440
- and stops being valid when you close the app.
432
+ only to find the conversation's id, and without keeping any line.
433
+ - **With OpenCode, the app runs a local OpenCode server** (`opencode serve`):
434
+ just one, from when you open the first OpenCode tab until five minutes after
435
+ you close the last one, or until you close the app. It listens on
436
+ `127.0.0.1`, on a random port, with a different password on every start, even
437
+ if your OpenCode configuration says otherwise. The app only asks it for the
438
+ sessions' status, pending permissions and questions, to create a session, to
439
+ answer a question and to stop a session: nothing about your configuration or
440
+ your accounts.
441
+
442
+ **Environment variables.** The CLIs inherit your environment as it is, with no
443
+ login variables added. Two exceptions: the app **removes**
444
+ `CLAUDE_CODE_CHILD_SESSION` —which turns off history saving— and tells you with
445
+ a banner when it does; and it **adds** one variable to the OpenCode server,
446
+ `OPENCODE_SERVER_PASSWORD`, with that per-start password. It isn't any account's
447
+ password, and it doesn't reach any tab: each OpenCode tab gets it on its command
448
+ line (`attach --password`), where other programs running as your user can see
449
+ it. It only works for that server and stops working when you close the app.
441
450
 
442
451
  **What the app writes:**
443
452
 
444
- - **In its own configuration directory:** the open tabs, the index cache, the
445
- notes, the archived sessions, the Antigravity status line script and, if you
446
- turn it on, the backup (or in the folder you choose). With remote access
447
- on, also the list of paired devices: a name, two dates and a hash of each
448
- one's credential — never the credential itself.
449
- - **In the temp folder:** the images you paste; each Antigravity tab's log,
450
- which contains your messages, readable only by you and deleted on the first
451
- start once it's more than 24 hours old; and the transcript of a conversation
452
- you continue in another CLI, deleted when you close that tab or after 24
453
- hours.
453
+ - **In its own configuration folder:** the open tabs, the index cache, the
454
+ notes, the archived sessions, unsent messages, the Antigravity status line
455
+ script and, if you turn it on, the backup (or in the folder you choose). With
456
+ remote access on, also the list of paired devices: a name, two dates and a
457
+ hash of each one's credential — never the credential itself.
458
+ - **In the temp folder:** the images and files you paste or attach; each
459
+ Antigravity tab's log, which contains your messages, readable only by you and
460
+ deleted on the first start after it's 24 hours old; and the transcript of a
461
+ conversation you continue with another CLI, deleted when you close that tab
462
+ or after 24 hours.
454
463
  - **In your projects, one thing only, and only if you confirm:** the shared
455
464
  memory. It's limited to `.agents/memory/`, to what's between its markers in
456
465
  `AGENTS.md` and `CLAUDE.md`, and to a few lines at the end of `.gitignore`.
457
466
 
458
- **The server listens only on `127.0.0.1`**, on an ephemeral port, with a random
459
- per-start token that the WebSocket and every HTTP route require, and it rejects
460
- requests whose `Origin` isn't its own. [Remote access](#remote-access-over-ssh-optional)
461
- doesn't change any of that: the port becomes a fixed one, still on `127.0.0.1`,
462
- and a paired device presents its own credential instead of the token. Reaching
463
- the port is left to SSH, which you turn on and control.
464
-
465
- **The git panel is read-only.** No commit, stage or push. With an agent editing
466
- files, a button that writes history is exactly the kind of thing where, later,
467
- nobody knows who pressed it.
467
+ </details>
468
468
 
469
469
  ![A file from the tree, previewed with syntax highlighting](https://raw.githubusercontent.com/cvelasquez/agent-workbench/main/assets/captura-archivos.png)
470
470
 
@@ -472,79 +472,95 @@ nobody knows who pressed it.
472
472
 
473
473
  ## Troubleshooting
474
474
 
475
- **The "claude" command wasn't found in the PATH**, followed by "Also works
476
- with: …" — the app didn't find any of the four CLIs in the `PATH` of the process
477
- running it. Check with `where claude` (or `which claude`), and the same with
478
- `codex`, `opencode` or `agy`. CLIs are looked up at startup: if you installed
479
- one while the app was open, restart it.
475
+ **"The "claude" command wasn't found in the PATH", followed by "Also works
476
+ with: …"** — the app didn't find any of the four CLIs. Check with
477
+ `where claude` on Windows (`which claude` elsewhere), and the same with `codex`,
478
+ `opencode` or `agy`. CLIs are looked up at startup: if you installed one while
479
+ the app was open, restart it.
480
480
 
481
- **The OpenCode history doesn't show up.** Look at the `History` line in the
482
- startup output. If it says that Node version doesn't include `node:sqlite`,
483
- update Node to 22.13 or later. If there's no such line, it didn't find the
484
- database: it's at `~/.local/share/opencode/opencode.db`, or wherever
485
- `OPENCODE_DB` or `XDG_DATA_HOME` points.
481
+ **OpenCode's history doesn't show up.** Look at the `History` line when the app
482
+ starts. If it says that Node version doesn't include `node:sqlite`, update Node
483
+ to 22.13 or later. If there's no such line, it didn't find the database: it's at
484
+ `~/.local/share/opencode/opencode.db`, or wherever `OPENCODE_DB` or
485
+ `XDG_DATA_HOME` points.
486
486
 
487
- **An OpenCode tab won't open and says "Couldn't start the OpenCode server".**
488
- The tab attaches to an `opencode serve` the app launches, and the reason comes
489
- after the colon. If it isn't clear, open `opencode` in a regular terminal: if it
490
- doesn't start there either, the problem is with that OpenCode installation.
487
+ **An OpenCode tab won't open: "Couldn't start the OpenCode server".** The reason
488
+ comes after the colon. If it isn't clear, open `opencode` in a regular terminal:
489
+ if it doesn't start there either, the problem is that OpenCode installation.
491
490
 
492
- **An OpenCode tab says the server closed.** The `opencode serve` process ended
493
- and the tab's terminal was left without a connection. **Relaunch**, in the same
494
- bar, starts another server and reattaches the tab to the same session.
491
+ **An OpenCode tab says the server closed.** **Relaunch**, in the same bar,
492
+ starts another server and reconnects the tab to the same session.
495
493
 
496
494
  **An Antigravity tab doesn't show its status or the meter.** Look at the
497
- `Status line` entry in the startup output, under the CLI: if it says it isn't
498
- configured, or that there's another one, follow the steps
499
- [above](#antigravity-cli-status-and-meter-optional). If you configured it and
500
- the CLI's terminal shows `Statusline Error`, `node` is most likely not in that
501
- session's `PATH`. Until the line publishes anything, the tab is treated as if
502
- you hadn't configured it.
495
+ `Status line` entry when the app starts, under the CLI. If it says it isn't
496
+ configured, or that there's another one, follow
497
+ [these steps](#antigravity-cli-status-and-context-meter). If you configured it
498
+ and the CLI's terminal shows `Statusline Error`, `node` is most likely not in
499
+ that session's `PATH`.
503
500
 
504
- **`node-pty` doesn't compile on install.** It's a native module. It usually
505
- downloads a prebuilt binary and nothing else is needed; if your combination of
506
- Node and platform doesn't have one, it has to be compiled:
501
+ **`node-pty` doesn't compile on install.** It's the terminal's native module. It
502
+ usually downloads a ready-made binary; if your Node and platform don't have one,
503
+ it has to be compiled:
507
504
 
508
505
  - **Windows:** Visual Studio Build Tools with the *Desktop development with C++*
509
- workload, and Python 3.
510
- `npm install --global windows-build-tools` is no longer maintained: install
511
- the Build Tools from the Visual Studio installer.
506
+ workload, and Python 3. (`windows-build-tools` from npm is no longer
507
+ maintained.)
512
508
  - **macOS:** `xcode-select --install`.
513
509
  - **Linux:** `build-essential` and `python3`.
514
510
 
515
511
  **The terminal stays blank.** The default renderer is canvas on purpose: with
516
- the WebGL addon the tab freezes on Windows 11 + Chrome even though the data
517
- arrives. If you want to try it anyway, add `?renderer=webgl` to the URL.
512
+ WebGL the tab freezes on Windows 11 + Chrome. To try WebGL anyway, add
513
+ `?renderer=webgl` to the address.
518
514
 
519
- **A banner says `CLAUDE_CODE_CHILD_SESSION` was removed.** It happens when you
520
- start the app from inside a Claude Code CLI session. That variable turns off
521
- history saving, and without history there's no conversation or meter. The app
522
- removes it and lets you know. In normal use —a regular terminal— it never shows
523
- up.
515
+ **A banner says `CLAUDE_CODE_CHILD_SESSION` was removed.** You started the app
516
+ from inside a Claude Code session. That variable turns off history saving, so
517
+ the app removes it and lets you know. From a regular terminal you'll never see
518
+ this.
524
519
 
525
520
  **From the other computer, `ssh` can't connect** (it times out or is refused).
526
- The SSH server isn't running on the computer that has the app, or something
527
- between the two blocks it. On Windows, check that the Wi‑Fi network is set to
528
- *Private* — the firewall rule the SSH server installs may not cover *Public*
529
- networks. And a work VPN that blocks the local network cuts this too while it's
530
- connected: there's nothing the app can do about that. If `ssh` rejects your
531
- user, a Microsoft or work account may need the name of your profile folder
532
- (`C:\Users\<name>`) and the account's password, not the PIN.
521
+ The SSH server isn't running on the computer with the app, or something between
522
+ the two blocks it. On Windows, check that the Wi‑Fi network is set to
523
+ *Private*: the SSH server's firewall rule may not cover *Public* networks. A
524
+ work VPN that blocks the local network also cuts it while it's connected. If
525
+ `ssh` rejects your user, a Microsoft or work account may need the name of your
526
+ user folder (`C:\Users\<name>`) and the account's password, not the PIN.
533
527
 
534
528
  **From the other computer, the page says "Origin not allowed"** (403). It was
535
- opened by the computer's name or IP. Open `http://localhost:<port>`, with the
536
- tunnel using the same port on both ends.
529
+ opened by the computer's name or IP address. Open `http://localhost:<port>`,
530
+ with the tunnel using the same port on both ends.
537
531
 
538
532
  **From the other computer, the page says the session token is missing** (401).
539
- That browser isn't paired, its access was revoked, or remote access is off.
540
- Pair it again from the ⇄ dialog.
533
+ That browser isn't paired, it was revoked, or remote access is off. Pair it
534
+ again from the ⇄ dialog.
541
535
 
542
536
  **The ⇄ dialog says the port is busy.** Another program —often a second Agent
543
- Workbench— already has it. Close it and restart, or choose another port.
537
+ Workbench— is using it. Close it and restart, or choose another port.
538
+
539
+ **The Changes panel says the folder isn't a git repository**, and it is one.
540
+ Make sure `git` is in the `PATH`. If the message is a different one, it's git's
541
+ own error, word for word.
542
+
543
+ ---
544
+
545
+ ## From source
546
+
547
+ To work on the app, or if you'd rather not install anything globally:
548
+
549
+ ```bash
550
+ corepack enable pnpm
551
+ pnpm install
552
+ pnpm dev # with hot reload
553
+ ```
554
+
555
+ ```bash
556
+ pnpm build # builds the interface once
557
+ pnpm start # serves the built interface
558
+ ```
544
559
 
545
- **The Changes panel says the folder isn't a git repository** and it is one.
546
- Make sure `git` is in the `PATH`. If the message is a different one, it's the
547
- error git returned, verbatim.
560
+ **One-click start on Windows:** `pnpm package` builds the interface and leaves
561
+ an **`Agent Workbench.cmd`** in the root. Double-click it: it installs whatever
562
+ is missing, builds if needed and opens the browser. It's a twenty-line text
563
+ file you can read before running it.
548
564
 
549
565
  ---
550
566
 
@@ -554,32 +570,24 @@ A pnpm monorepo, TypeScript throughout.
554
570
 
555
571
  ```
556
572
  packages/
557
- server/ Node, Express, ws, node-pty, chokidar — serves the interface and hosts the ptys
573
+ server/ Node, Express, ws, node-pty, chokidar — serves the interface and runs the terminals
558
574
  src/agents/ one adapter per CLI: the only part of the server that knows each one
559
575
  web/ Vite, React, xterm.js, highlight.js
560
576
  shared/ the protocol types, with no `any` at the edges
561
- android/ the Android app, in Kotlin: the SSH tunnel, a viewer for the interface, the notifications
577
+ android/ the Android app, in Kotlin: the SSH tunnel, the interface, the notifications
562
578
  ```
563
579
 
564
- A single process serves the interface and the WebSocket on the same port: with
565
- a single origin, the `Origin` check and the token work the same in development
566
- and in production, with no exceptions that nobody remembers to remove later.
567
-
568
- The architectural decision that shapes the rest: **the ptys live in a server
569
- registry and the WebSocket is just transport.** If the process died with the
570
- socket, an accidental `Ctrl+R` would wipe out the working session. Each terminal
571
- keeps a buffer of its recent output to repaint the screen when the client comes
572
- back.
573
-
574
- The other one: **the generic server doesn't name any CLI.** What it knows about
575
- each one —where it stores things, how it's launched, what is read and what is
576
- never opened— lives in its adapter, and the interface draws each control based
577
- on what that CLI declares.
578
-
579
- [`ARCHITECTURE.md`](ARCHITECTURE.md) has the hard rules, the code map, what is
580
- read from each CLI and what is never opened. [`CONTRIBUTING.md`](CONTRIBUTING.md)
581
- covers how to work on the repository, [`SECURITY.md`](SECURITY.md) how to report
582
- a vulnerability, [`PRIVACY.md`](PRIVACY.md) what is collected (nothing), and
580
+ Two decisions shape the rest. **The terminals live on the server and the
581
+ browser is just a window onto them**: closing or reloading the page doesn't
582
+ kill a working session. And **the generic server doesn't name any CLI**: what it
583
+ knows about each one —where it keeps things, how it's launched, what is read
584
+ and what is never opened— lives in that CLI's adapter, and the interface shows
585
+ each control based on what that CLI declares.
586
+
587
+ [`ARCHITECTURE.md`](ARCHITECTURE.md) has the hard rules and the code map,
588
+ [`CONTRIBUTING.md`](CONTRIBUTING.md) how to work on the repository,
589
+ [`SECURITY.md`](SECURITY.md) how to report a vulnerability,
590
+ [`PRIVACY.md`](PRIVACY.md) what is collected (nothing), and
583
591
  [`CHANGELOG.md`](CHANGELOG.md) what changed in each version. Code comments are
584
592
  in Spanish. The screenshots in this README come from `pnpm demo:shots`, on
585
593
  made-up data.