@gaiventures/pilotfish-mcp-managed 0.1.36 → 0.1.38
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 +31 -33
- package/cli.js +1 -1
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -2,23 +2,29 @@
|
|
|
2
2
|
|
|
3
3
|
[](https://www.npmjs.com/package/@gaiventures/pilotfish-mcp-managed)
|
|
4
4
|
|
|
5
|
-
|
|
5
|
+
An MCP server that does whole browser tasks for your AI agent. Your agent says what it wants done ("download last
|
|
6
|
+
month's invoices from our AWS billing console", "pull the risk factors from Apple's latest 10-K on SEC EDGAR").
|
|
7
|
+
[pilotfish](https://pilotfish.gai.ventures) does it in its cloud browser, signed in with your saved logins, and
|
|
8
|
+
returns the answer and any files.
|
|
6
9
|
|
|
7
|
-
|
|
8
|
-
|
|
9
|
-
|
|
10
|
+
For people and teams using Claude Code, Claude.ai, Claude Desktop, Codex CLI, Cursor, VS Code or another
|
|
11
|
+
[MCP](https://modelcontextprotocol.io) client, and for products that run browser tasks for their own users. It is a
|
|
12
|
+
paid service: you need a pilotfish account ([sign up with Google](https://pilotfish.gai.ventures/signup)), and tasks
|
|
13
|
+
are paid from your organization's credits.
|
|
10
14
|
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
|
|
15
|
+
```sh
|
|
16
|
+
claude mcp add --transport http pilotfish https://pilotfish-be.gai.ventures/mcp
|
|
17
|
+
```
|
|
14
18
|
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
19
|
+
Then run `/mcp` in Claude Code and sign in. Any client that takes a URL connects to
|
|
20
|
+
`https://pilotfish-be.gai.ventures/mcp`, by signing in or with an API key. This npm package is a small relay for clients that can only
|
|
21
|
+
start a local (stdio) server: `npx -y @gaiventures/pilotfish-mcp-managed` (setup under "Connect" below).
|
|
18
22
|
|
|
19
23
|
Want a browser on your own computer for free, driven by your agent's own model, with no account? Use the open
|
|
20
24
|
[`@gaiventures/pilotfish-mcp`](https://www.npmjs.com/package/@gaiventures/pilotfish-mcp) instead.
|
|
21
25
|
|
|
26
|
+
[](https://mcprush.com/to/pilotfish-managed-mcp?ref=badge)
|
|
27
|
+
|
|
22
28
|
## Why managed
|
|
23
29
|
|
|
24
30
|
- **One call per task.** pilotfish plans and runs the task on its own models and checks the answer. Your agent spends
|
|
@@ -182,7 +188,10 @@ make one in the app's **Org settings → API keys** and pick the project **Perso
|
|
|
182
188
|
Personal project, with your own saved logins, and only you can see or manage it. Admins (the primary admin included)
|
|
183
189
|
can always do this; a member can once an admin turns on **Allow personal API keys** on the member's page (turning it
|
|
184
190
|
off revokes their keys at once), and then finds their keys on their Org settings page. Admins see each person's key
|
|
185
|
-
usage, never the keys.
|
|
191
|
+
usage, never the keys. Every key, personal ones included, can have a spending limit of its own: admins set it on the
|
|
192
|
+
key's page (a member sees the limit and how much of it is used, and asks an admin to change it). At the limit, new
|
|
193
|
+
tasks with that key are refused and running ones pause until it is raised; `get_usage` with a personal key says what
|
|
194
|
+
is left of it. Most people need no key at all: Claude Code, the Claude app, Cursor and VS Code sign in, and
|
|
186
195
|
a client that starts this package signs in once with `login` (see **Sign in a terminal**).
|
|
187
196
|
|
|
188
197
|
## Projects
|
|
@@ -397,30 +406,19 @@ machine running it stopped or was replaced says `PAUSED` (`machine_lost`, `machi
|
|
|
397
406
|
does: `follow_up` with the text "Continue" picks it up from where it left off, in a new browser. A job whose machine
|
|
398
407
|
stopped before it did anything on a site starts again by itself, once.
|
|
399
408
|
|
|
400
|
-
##
|
|
409
|
+
## Writing tasks
|
|
401
410
|
|
|
402
411
|
Send the whole web task in one `browser_task`, however many sites and steps it has: "find the cheapest return flight
|
|
403
|
-
on MakeMyTrip, then book a hotel near the beach for those dates on Booking.com"
|
|
404
|
-
|
|
405
|
-
|
|
406
|
-
|
|
407
|
-
|
|
408
|
-
|
|
409
|
-
|
|
410
|
-
|
|
411
|
-
|
|
412
|
-
|
|
413
|
-
pilotfish searches the web first and answers from what it read, every fact linked to its page, in seconds and with no
|
|
414
|
-
browser. When the answer needs a site (a fare for a date, a table to book, your own account), the browser starts on
|
|
415
|
-
the site the search found; a price or availability for particular dates is always read on the site itself.
|
|
416
|
-
|
|
417
|
-
## When the person says more mid-task
|
|
418
|
-
|
|
419
|
-
Your agent reads what the person types only between tool calls, so `browser_task`, `follow_up` and `get_job` return
|
|
420
|
-
within two minutes by default (50 s for a client that takes no progress updates, since most give up on a call after
|
|
421
|
-
60 s) while the task carries on. Anything the person says meanwhile (a correction, a detail,
|
|
422
|
-
the answer to the task's question) goes to the task with `message_job`, in their words, and the agent keeps waiting
|
|
423
|
-
with `get_job`.
|
|
412
|
+
on MakeMyTrip, then book a hotel near the beach for those dates on Booking.com". pilotfish splits it into parts, asks
|
|
413
|
+
for every sign-in it needs at the start, carries what each site gives to the next and checks the end result. Say in
|
|
414
|
+
the task what may be completed without asking ("pay with my saved card", "go ahead and submit"); anything else that
|
|
415
|
+
is final waits for a yes.
|
|
416
|
+
|
|
417
|
+
A task that names no site ("research Postgres vs MySQL for a startup") may be answered from a web search, every fact
|
|
418
|
+
linked to its page, with no browser. A price or availability for particular dates is always read on the site itself.
|
|
419
|
+
|
|
420
|
+
Your agent gets the rest from the tools' own descriptions: when to split a task itself, and how to pass on what the
|
|
421
|
+
person says while a task runs (`message_job`).
|
|
424
422
|
|
|
425
423
|
## When a job needs a person
|
|
426
424
|
|
package/cli.js
CHANGED
|
@@ -15,7 +15,7 @@ import { Client } from "@modelcontextprotocol/sdk/client/index.js";
|
|
|
15
15
|
import { isJSONRPCRequest, isJSONRPCResultResponse } from "@modelcontextprotocol/sdk/types.js";
|
|
16
16
|
|
|
17
17
|
// packages/mcp-managed/dist/version.js
|
|
18
|
-
var VERSION = true ? "0.1.
|
|
18
|
+
var VERSION = true ? "0.1.38" : "0.0.0-dev";
|
|
19
19
|
|
|
20
20
|
// packages/mcp-managed/dist/bridge.js
|
|
21
21
|
var DEFAULT_URL = "https://pilotfish-be.gai.ventures";
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@gaiventures/pilotfish-mcp-managed",
|
|
3
|
-
"version": "0.1.
|
|
3
|
+
"version": "0.1.38",
|
|
4
4
|
"description": "Whole browser tasks done in pilotfish's cloud for any MCP agent, signed in with your saved logins. mcprush.com-37c48d491684ec6011385ea5",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"type": "module",
|