@ricardochen1996/browser-use-mcp 0.0.0-stage → 0.1.0

package/docs/clients.md

@@ -0,0 +1,425 @@
+# Client wiring
+
+`browser-use-mcp` speaks the Model Context Protocol. This page has copy-pasteable configuration for
+the clients people actually use: **[Claude Code](#claude-code)**, **[Claude Desktop](#claude-desktop)**,
+**[Cursor](#cursor)**, **[Codex](#codex)**, **[DSH](#dsh)**, and a plain
+**[`mcpServers` JSON](#generic-mcpservers-json)** any MCP host understands. Every stdio snippet
+starts `node` with the server's absolute path and runs from any working directory.
+
+## Before you paste: three things
+
+1. **Use your own absolute path.** The snippets below use this checkout's path,
+   `/Users/ricardo.chen/PersonalGit/browser-use-mcp`. Replace it with wherever you cloned the
+   repository. Do not rely on a relative path: GUI applications do not start in the directory you
+   ran a shell command from.
+2. **Make sure `node` is the Node 22+ you expect.** The server needs Node `^22.19.0 || >=24.0.0`.
+   A GUI-launched client often has a shorter `PATH` than your shell. If it cannot find `node`, run
+   `which node` in your shell and put that absolute path in `command` instead — for example
+   `/opt/homebrew/bin/node`.
+3. **stdout is the protocol; logs are stderr.** While serving MCP, stdout carries JSON-RPC frames
+   and every log line goes to stderr. When a client shows nothing, run the exact command in a
+   terminal: the JSON-RPC traffic appears on stdout and `browser-use-mcp …` log lines on stderr,
+   which is where startup and engine-install problems are explained.
+
+The package is **not published to npm yet**. Install from the checkout:
+
+```bash
+cd /Users/ricardo.chen/PersonalGit/browser-use-mcp
+npm install
+```
+
+That is enough — the server itself is plain ESM. The Python side (the engine and `browser-harness`)
+is provisioned automatically on first use, or ahead of time:
+
+```bash
+node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs install
+node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs doctor
+```
+
+If you prefer a command name over an absolute path, `npm link` in the checkout puts
+`browser-use-mcp` on your `PATH`; GUI clients may still not see that `PATH`, so the absolute path is
+the reliable form. Every snippet below is **stdio** unless it says otherwise; the server's default
+transport is stdio, one browser per server process.
+
+Two defaults decide what the agent sees:
+
+- **Transport stdio** — the client starts its own server process, and only that client uses it.
+- **Mode attach** — the server drives the Chrome you already use, so the agent sees your logins and
+  tabs. Enable remote debugging once in that Chrome (`chrome://inspect/#remote-debugging` → allow);
+  on macOS the application running the server also needs Full Disk Access. If you would rather give
+  the agent a dedicated browser, add `BROWSER_USE_MCP_MODE=launch` (and `BROWSER_USE_MCP_HEADLESS=1`
+  for no window). `BROWSER_USE_MCP_PROFILE` and `BROWSER_USE_MCP_HEADLESS` only mean something in
+  launch mode — the snippets below set `launch` where they use them.
+
+---
+
+## Claude Code
+
+Two equivalent ways to wire it.
+
+### CLI
+
+```bash
+# Attach mode (the default): use the Chrome you browse with.
+claude mcp add --scope user browser-use \
+  -- node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs
+
+# Launch mode: a dedicated headless browser with its own profile.
+claude mcp add --scope user browser-use \
+  --env BROWSER_USE_MCP_MODE=launch \
+  --env BROWSER_USE_MCP_HEADLESS=1 \
+  --env BROWSER_USE_MCP_PROFILE=claude-code \
+  -- node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs
+```
+
+- `--scope user` makes it available in every project. Use `--scope project` to write a `.mcp.json`
+  in the current repository instead, or omit it for local-only.
+- Add `--env TYPESAFE_API_KEY=…` (or the `TEXT_MODEL_*` group) to enable `browser_goal` and
+  `browser_act` intents; see [Jev credentials](#jev-credentials).
+- Check it with `claude mcp list`, then ask Claude to call `browser_doctor`.
+
+### Project `.mcp.json`
+
+```json
+{
+  "mcpServers": {
+    "browser-use": {
+      "command": "node",
+      "args": ["/Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs"],
+      "env": {
+        "BROWSER_USE_MCP_MODE": "launch",
+        "BROWSER_USE_MCP_HEADLESS": "1",
+        "BROWSER_USE_MCP_PROFILE": "claude-code"
+      }
+    }
+  }
+}
+```
+
+---
+
+## Claude Desktop
+
+Edit `~/Library/Application Support/Claude/claude_desktop_config.json` on macOS, or
+`%APPDATA%\Claude\claude_desktop_config.json` on Windows. The file is plain JSON (no comments, no
+trailing commas):
+
+```json
+{
+  "mcpServers": {
+    "browser-use": {
+      "command": "node",
+      "args": ["/Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs"],
+      "env": {
+        "BROWSER_USE_MCP_MODE": "launch",
+        "BROWSER_USE_MCP_PROFILE": "claude-desktop"
+      }
+    }
+  }
+}
+```
+
+Fully quit and restart Claude Desktop after saving. In launch mode the browser window appears when
+the model calls `browser_open`, not at startup; in attach mode it is already the Chrome you use.
+`BROWSER_USE_MCP_MODE=launch` is what makes this client use a dedicated profile — give each client
+its own `BROWSER_USE_MCP_PROFILE` if several clients will launch at once (two servers on one profile
+are refused by design).
+
+---
+
+## Cursor
+
+Project configuration goes in `.cursor/mcp.json`; to make it available everywhere, use
+`~/.cursor/mcp.json`:
+
+```json
+{
+  "mcpServers": {
+    "browser-use": {
+      "command": "node",
+      "args": ["/Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs"],
+      "env": {
+        "BROWSER_USE_MCP_MODE": "launch",
+        "BROWSER_USE_MCP_HEADLESS": "1",
+        "BROWSER_USE_MCP_PROFILE": "cursor"
+      }
+    }
+  }
+}
+```
+
+Reload the window (or restart Cursor) after editing. The tools appear in **Settings → MCP** once the
+server answers an `initialize`.
+
+---
+
+## Codex
+
+Add the server to `~/.codex/config.toml`:
+
+```toml
+[mcp_servers.browser-use]
+command = "node"
+args = ["/Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs"]
+startup_timeout_sec = 120
+
+[mcp_servers.browser-use.env]
+BROWSER_USE_MCP_MODE = "launch"
+BROWSER_USE_MCP_HEADLESS = "1"
+BROWSER_USE_MCP_PROFILE = "codex"
+```
+
+`startup_timeout_sec` is optional; a first run that has to provision the Python environment can take
+a minute or two, so raise it if your Codex build gives up before the server answers. Restart Codex
+after saving, then check that `browser_doctor` is in its tool list.
+
+---
+
+## DSH
+
+DSH loads MCP servers through its `@deepseek-ai/dsh-mcp-client` plugin, configured in the profile's
+`cordis.patch.yml` (`$DSH_HOME/profiles/<profile>/cordis.patch.yml`; `DSH_HOME` defaults to
+`~/.dsh`, and the standard Web UI profile is `web`). The row is namespaced: its tools appear as
+`mcp__browser-use__browser_open` and so on.
+
+**stdio** — DSH starts the server itself:
+
+```yaml
+- insert:
+    - id: browser-use-mcp
+      name: '@deepseek-ai/dsh-mcp-client'
+      config:
+        serverName: browser-use
+        transport: stdio
+        command: node
+        args:
+          - /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs
+        env:
+          BROWSER_USE_MCP_MODE: launch
+          BROWSER_USE_MCP_HEADLESS: '1'
+          BROWSER_USE_MCP_PROFILE: dsh
+        failOnStartupError: true
+```
+
+**Streamable HTTP** — point at a server you started yourself (see
+[one shared HTTP server](#one-shared-http-server)):
+
+```yaml
+- insert:
+    - id: browser-use-mcp
+      name: '@deepseek-ai/dsh-mcp-client'
+      config:
+        serverName: browser-use
+        transport: streamable-http
+        url: http://127.0.0.1:7331/mcp
+        failOnStartupError: true
+```
+
+A DSH detail worth knowing: the stdio transport merges the configured `env` over a **scrubbed**
+ambient environment, and ambient names matching `KEY`, `PASSWORD`, `SECRET` or `TOKEN` are dropped.
+So Jev credentials must be listed explicitly in `env` — they do not leak in from the shell:
+
+```yaml
+        # Jev credentials must be repeated in env: ambient names matching KEY/TOKEN are scrubbed.
+        env:
+          BROWSER_USE_MCP_MODE: launch
+          BROWSER_USE_MCP_HEADLESS: '1'
+          TYPESAFE_API_KEY: !!js process.env.TYPESAFE_API_KEY
+```
+
+Restart the harness if the row is new. Both variants accept the same other `config` fields as any
+`dsh-mcp-client` entry (`reconnect`, `toolCallTimeoutMs`, …).
+
+---
+
+## Generic `mcpServers` JSON
+
+Most MCP hosts read a JSON object with one entry per server. This is the same shape as the Claude
+Desktop and Cursor snippets, with the most common knobs shown:
+
+```json
+{
+  "mcpServers": {
+    "browser-use": {
+      "command": "node",
+      "args": [
+        "/Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs",
+        "--mode", "launch",
+        "--headless",
+        "--profile", "my-agent"
+      ],
+      "env": {
+        "BROWSER_USE_MCP_HEADLESS": "1",
+        "BROWSER_USE_MCP_PROFILE": "my-agent",
+        "BROWSER_USE_MCP_HOME": "/Users/me/.browser-use-mcp"
+      }
+    }
+  }
+}
+```
+
+Server flags go in `args`; environment variables go in `env`. `--headless` and
+`BROWSER_USE_MCP_HEADLESS=1` mean the same thing — either spelling is fine, and a flag wins if both
+are present. The full list is in the [configuration reference](config.md).
+
+Mutually equivalent spellings for the same settings:
+
+| Setting | Flag | Env |
+| --- | --- | --- |
+| Headless | `--headless` | `BROWSER_USE_MCP_HEADLESS=1` |
+| Profile | `--profile my-agent` | `BROWSER_USE_MCP_PROFILE=my-agent` |
+| State root | `--home /path` | `BROWSER_USE_MCP_HOME=/path` |
+| No `browser_eval` | `--no-eval` | `BROWSER_USE_MCP_NO_EVAL=1` |
+| Attach to a debug browser | `--mode attach --cdp-endpoint http://127.0.0.1:9333` | `BROWSER_USE_MCP_MODE=attach`, `BROWSER_USE_MCP_CDP_ENDPOINT=http://127.0.0.1:9333` |
+
+### Environment examples
+
+Headless, separate profile, explicit state root:
+
+```json
+"env": {
+  "BROWSER_USE_MCP_HEADLESS": "1",
+  "BROWSER_USE_MCP_PROFILE": "agent-a",
+  "BROWSER_USE_MCP_HOME": "/Users/me/.browser-use-mcp"
+}
+```
+
+Jev credentials belong in `env`, never in `args` (arguments are visible in process listings):
+see the next section.
+
+---
+
+## Jev credentials
+
+`browser_goal` and `browser_act`'s `intent` are optional; the rest of the tool surface never spends
+model quota. Two ways to make the policy work:
+
+**TypeSafe decider** (default) — Jev answers the policy questions:
+
+```json
+"env": {
+  "TYPESAFE_API_KEY": "your-key",
+  "TYPESAFE_BASE_URL": "https://…",
+  "TYPESAFE_MODEL": "…"
+}
+```
+
+Only `TYPESAFE_API_KEY` is required; the other two have engine defaults. `TEXT_MODEL_*` is still
+wanted if the policy should write `TYPE_TEXT` values.
+
+**Chat decider** — a text model answers instead:
+
+```json
+"env": {
+  "JEV_DECIDER": "chat",
+  "TEXT_MODEL_BASE_URL": "https://your-endpoint/v1",
+  "TEXT_MODEL": "your-model",
+  "TEXT_MODEL_API_KEY": "your-key"
+}
+```
+
+`--jev-decider chat` in `args` is equivalent to `JEV_DECIDER=chat`. To keep the surface entirely
+policy-free, pass `--no-jev` (or `BROWSER_USE_MCP_NO_JEV=1`).
+
+The doctor reports only `set`/`unset` for keys, never a value, and a missing key is a refusal that
+names the variable, not a crash.
+
+---
+
+## One shared HTTP server
+
+Running a long-lived HTTP endpoint is how several agents share one installed server instead of each
+starting a stdio process. Each MCP session still gets its own browser session, sidecar and tab, so
+two agents never click on one page by accident.
+
+Start it once:
+
+```bash
+# Default: every session drives the Chrome you already use (attach, shared by default).
+node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs \
+  --transport http --host 127.0.0.1 --port 7331 --path /mcp
+
+# Or: every session gets a tab in one dedicated browser this server launches.
+node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs \
+  --transport http --host 127.0.0.1 --port 7331 --path /mcp \
+  --mode launch --headless
+```
+
+The endpoint is then `http://127.0.0.1:7331/mcp`. Flags given here apply to every session; clients
+only supply the URL.
+
+**Persistent per-agent profiles (launch mode).** An anonymous session's profile is deleted when the
+session ends. To keep logins across reconnects, the client names its profile on the initialize
+request — in the URL, `http://127.0.0.1:7331/mcp?profile=claude`, or with the header
+`X-Browser-Use-Profile: claude`. Different names run at once on different profiles; two live
+sessions with the same name are refused (the profile is one browser's at a time). The snippets below
+show the URL form.
+
+- **Claude Code**
+
+  ```bash
+  claude mcp add --transport http --scope user browser-use \
+    "http://127.0.0.1:7331/mcp?profile=claude"
+  ```
+
+- **Claude Code project config**
+
+  ```json
+  {
+    "mcpServers": {
+      "browser-use": { "type": "http", "url": "http://127.0.0.1:7331/mcp?profile=claude" }
+    }
+  }
+  ```
+
+- **Cursor** (`~/.cursor/mcp.json`)
+
+  ```json
+  {
+    "mcpServers": {
+      "browser-use": { "url": "http://127.0.0.1:7331/mcp?profile=cursor" }
+    }
+  }
+  ```
+
+- **Codex** (`~/.codex/config.toml`)
+
+  ```toml
+  [mcp_servers.browser-use]
+  url = "http://127.0.0.1:7331/mcp"
+  ```
+
+- **DSH** — use the `streamable-http` row [above](#dsh).
+
+- **Any client that only says "URL"** — use `http://127.0.0.1:7331/mcp`, or add
+  `?profile=<name>` for a persistent per-agent profile in launch mode.
+
+Notes:
+
+- There is **no authentication**. Keep the bind on `127.0.0.1`; if it must be reachable from
+  elsewhere, put it behind a tunnel or reverse proxy you control.
+- An idle session's browser is closed after 15 minutes; raise or lower that with
+  `BROWSER_USE_MCP_HTTP_IDLE_MS` (milliseconds, positive) in the server's environment.
+- A client that disappears without terminating its session is cleaned up by the same idle sweep; an
+  explicit terminate cleans up at once.
+- In launch mode browsers are launched per session with a profile directory under the state root:
+  anonymous ones are removed when the session ends, a `?profile=<name>` one is kept. In attach mode
+  (the default) every session drives the same already-running browser, shared by default — see
+  `--attach-sharing` in the [configuration reference](config.md).
+
+---
+
+## If a client shows no tools
+
+| Symptom | Likely cause | Fix |
+| --- | --- | --- |
+| Client says the server exited immediately | `node` not found, or a bad flag value | Run the same command in a terminal and read stderr; a configuration error exits with code `2` and names the field. |
+| Server starts but no tools appear | Wrong config file, or the client was not restarted | Use the client's MCP list/log view; restart the client after editing its config. |
+| First call takes minutes, then works | First run is provisioning the Python environment | Wait, or run `node /Users/ricardo.chen/PersonalGit/browser-use-mcp/bin/browser-use-mcp.mjs install` first. Behind a proxy, set `HTTPS_PROXY`. |
+| `browser_doctor` reports `Status: ready` but calls fail | The profile is held by another server, or the browser process died | Read the failure text — a profile collision names `--profile`; a dead browser is rebuilt on the next `browser_open`. |
+| Profile lock error naming `--profile` | Two servers on one browser profile | Stop the other server or give this one `--profile <other-name>`. |
+| Attach mode cannot find your Chrome (macOS) | Full Disk Access missing | System Settings → Privacy & Security → Full Disk Access, add the host app, restart it. |
+| Chrome shows "Allow remote debugging?" | A second CDP client is running | Approve the prompt once, never retry; stop the extra client. The session's own tools share one connection. |
+
+For the design behind the freshness rule and the one-connection rule, read
+[architecture.md](architecture.md). For every flag and variable, read [config.md](config.md).