Pilot Browse MCP
An MCP toolset that lets AI agents drive your real Chrome, learn websites, and reuse the resulting manuals.
- Source repo
- EndymionLee/PilotBrowseMCP
- Stars
- ★ 23
- Last updated
- 3mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 37/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers wiring Claude Code, Pi or Codex to a real, logged-in Chrome session through an MCP client.
- Not a fit if
- Users who only want a hosted browser and won't self-host
- Maintenance
- Commits in the last 6 monthslast commit 2026-08-01
- Source review
- 37/100 · Major gaps 2 safety controls not found
What does this agent do, and when should you use it?
Pilot Browse MCP is a three-part browser runtime: a Node.js MCP server, a Chrome extension, and the real browser they drive over WebSocket :9456. An agent speaks MCP stdio (JSON-RPC) and gets 62 tools for tab management, content extraction, DOM actions, network interception, file saving, workflow recording, script execution, SQL injection checks and JS reverse engineering. While exploring a site it captures APIs and binds them with browser flows into reusable capabilities stored under website-manuals/, with pages/, navigation/, workflows/ (.pab scripts) and apis/endpoints/. Those manuals can be reused by later agents, or executed without an LLM via .pab scripts or generated Python. The project is MIT-licensed and self-hosted: you build both components and load the extension yourself.
The MCP server receives JSON-RPC over stdio and forwards work through WebSocket :9456 to the Chrome extension, which calls browser APIs. Page tools include browser_get_markdown (Readability + Turndown), browser_get_text, browser_get_html, browser_find, browser_query (penetrates Shadow DOM), browser_evaluate, and browser_extract_article/table/links/images. Actions include browser_click (composed:true for Shadow DOM), browser_type, browser_scroll, browser_wait and browser_wait_for_element. browser_save_content and browser_save_xpath write extracted content straight to disk, bypassing the LLM to save tokens. Network tools start/stop monitoring, search and inspect cached requests, wait for a matching request after an action, replay with query/header/body overrides, export as curl/fetch/Python/HAR, and analyze a site's API structure. Workflow tools turn popup recordings and marked elements into entries in website-manuals/, and can generate and execute MCP automation scripts.
- Developers wiring Claude Code, Pi or Codex to a real, logged-in Chrome session through an MCP client.
- Data collectors who want structured output without paying LLM tokens per page, using browser_save_content to write files directly.
- Teams that want to accumulate a shared manual of how a site works so later agents ramp up faster.
- Engineers reverse-engineering APIs, using js_analyze, js_find_function, js_trace_request and network replay.
- Security testers replaying requests in browser context to check for SQL injection.
- Operators running bulk automation without an LLM via .pab scripts or generated Python.
How do you install or deploy this agent?
You need Node.js, npm and desktop Chrome. Build both components, load the extension unpacked, then start the server:
# Build
cd server && npm install && npm run build
cd extension && npm install && npm run build
# Load extension
# chrome://extensions/ -> Developer mode -> Load unpacked -> extension/dist/
# Start server
cd server && node dist/index.jsPoint your MCP client's args at the actual server/dist/index.js path:
{
"mcpServers": {
"browser-mcp": {
"command": "node",
"args": ["/path/to/server/dist/index.js"]
}
}
}Keep the server running and the extension loaded; they talk over WebSocket :9456.
How do you use this agent?
Start the MCP server, load the extension, then call tools from any MCP stdio client. agent-examples/ contains ready-to-use agent workspaces that auto-load MCP config, skills and project prompts; update the MCP path to your own. A typical session: browser_start_network_monitor, locate the page with the tab/page tools, explore structure with browser_inspect_page and browser_evaluate, persist findings with workflow_generate into website-manuals/, then reuse via workflow_execute_script or a .pab script. Available skills live in scripts/Skill.
What are this agent's strengths and limitations?
- Broad tool surface: the README documents 62 MCP tools across page, action, network, tab, recording, data, security and JS-reverse categories.
- browser_save_content and browser_save_xpath write pages to disk without routing content through the LLM, cutting token spend.
- website-manuals/ organizes learned capabilities into pages, navigation, workflows and apis for later reuse.
- The PAB DSL supports if/for/fn and runs from the popup, so automation can run with no LLM at all.
- The network toolkit covers monitoring, search, replay with overrides, response overrides and export to curl/fetch/Python/HAR.
- Multi-step setup: install and build both server and extension, load dist unpacked in chrome://extensions, then configure an MCP client.
- Tied to desktop Chrome through the extension model; no Firefox, Safari or headless path is documented.
- Server and extension communicate over a fixed WebSocket port (:9456), adding a connection to keep healthy.
- Sensitive tools (cookies, localStorage, screenshot) need explicit permissions, and the README does not detail the full permission model.
- No package-manager install or one-shot upgrade path; both sides must be rebuilt when updating.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Pilot Browse MCP This agent | 37 · Major gaps | MCP serverFree + model costs | ★ 23 | 3mo ago | TypeScript | — |
| OpenTabs | 48 · Major gaps | CLIFree + model costs | ★ 1k | 3mo ago | TypeScript | Claude Code |
| Uni-CLI | 83 · Good | CLIFree | ★ 229 | 5d ago | TypeScript | ChatGPT · Codex · Claude Code |
| Interceptor | 69 · Some gaps | CLIFree | ★ 638 | 1d ago | TypeScript | Codex · Claude Code · Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
A permission store and popup toggles (cookies/local_storage/screenshot/get_html/sql_injection) show least-privilege awareness, but all 62 tools are registered by default, browser_evaluate runs arbitrary JS, and js_hook_inject injects into MAIN world and hooks CryptoJS inputs/outputs; granularity is coarse, so least_privilege is 1. User confirmation is only a 'requires permission' note on sensitive tools with no per-action confirmation flow, user_confirmation 1. The README architecture diagram documents the MCP→WebSocket:9456→extension→browser path but not data destination/retention, data_flow_transparency 1. Sensitive data (cookies, localStorage, screenshots, hooked crypto I/O) is gated but has no redaction/encryption/retention policy, sensitive_data_handling 1. No lockfile audit or dependency inventory evidence, dependency_security 0. External effects (clicks, typing, network replay, active SQLi scanning) have no sandbox or impact-scope statement, external_effects 1. No rollback/undo for network overrides, script execution, or file writes, rollback 0. MIT license and copyright notice exist but the author is unnamed ('Copyright (c) 2026'), source_attribution 1.
SKILL.md is duplicated identically across scripts/Skill and three agent-examples, which is self-consistent, but README claims 62 tools while SKILL references workflow_validate_manual that is absent from the README tool table, self_consistency 1. Install steps give npm install/build and MCP config; dependency availability is inferable but no version pinning evidence, dependency_availability 1. Code has error handling and [FATAL] output and pab_run returns error codes, but several empty catches (SQLi listener, js_collect fetch) leave failures silent, failure_messages 1.
README targets Claude Code/Pi/Codex agents and ships agent-examples workspaces, audience_and_scenarios 2. Capability boundaries are implied by the tool table and phases but unsupported sites/operations and legal limits of SQLi scanning are unstated, capability_boundaries 1. Trigger precision relies on SKILL description and tool names with no explicit trigger/exclusion rules, trigger_precision 1. Environment requires Chrome extension + Node server + localhost:9456 with no cross-platform/version requirements stated, environment_fit 1.
README has an architecture diagram, categorized tool table, and output directory structure, information_architecture 2. Install steps are complete (build, load extension, start server, MCP config), install_notes 2. Naming is mostly consistent between SKILL and README but workflow_validate_manual is missing from the tool table, naming_stability 1. Demo GIFs and example directories exist but no FAQ, examples_and_faq 1. No known-limitations section, known_limitations 0. Full MIT license, license 2. package.json version 0.1.0 but no CHANGELOG, versioning_changelog 0. No maintainer/contact/issue process, maintenance_responsibility 1.
Output is a structured website-manuals tree (pages/navigation/workflows/apis) plus .pab and Python script generation, output_usability 2. Manual reuse and token savings are the stated value, marginal_value 2. Cost-benefit lacks quantified evidence (only an anecdote about crawling 4000 novels on a NAS) and 62 tools plus multi-phase flow may be heavy for simple tasks, cost_benefit 1.
README demos and performance claims (4000 novels in 2 days) have no logs or reproduction steps, claim_traceability 1. SKILL and README corroborate the output directory and phases, but tool count and tool names disagree, cross_source_corroboration 1. Code and docs mostly separate facts from inference, but README states capabilities like 'automatically discovers APIs' as established fact without evidence, fact_inference_separation 1.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- browser_evaluate runs arbitrary page JS and js_hook_inject hooks CryptoJS and logs crypto inputs/outputs, which may capture sensitive credentials; use only on trusted sites.
- Active SQL injection scanning (sql_injection_scan/request) can generate offensive traffic against third-party sites; confirm authorization and legal compliance.
- Network replay and response overrides have no rollback and may alter site state or data.
- Cookies/localStorage/screenshot permissions are coarse toggles with no per-action confirmation or redaction policy.
- README claims 62 tools but the tool table and SKILL references are not fully consistent, and capability claims lack reproducible evidence.
- No CHANGELOG, no known-limitations section, and no maintainer contact; long-term maintenance and update path are unclear.
FAQ
Does it add model costs?
Is Chrome required?
How does it save tokens?
Which MCP clients does it work with?
Is Pilot Browse MCP open source?
Is Pilot Browse MCP safe to use?
Is Pilot Browse MCP still maintained?
Who makes Pilot Browse MCP?
Thanks — we use this to improve the page.