Auto Browser

Give your AI agent a real browser, with a human in the loop.

Auto Browser is an MCP-native browser control plane for authorized workflows. It gives MCP clients, LLM agents, and operators a shared Playwright browser with human takeover, reusable auth profiles, approvals, audit trails, and local-first deployment.

Works with:

• Claude Code, Codex (CLI, IDE extension, and app), Google Antigravity, Cursor, and VS Code, connected directly over HTTP

• Claude Desktop, through the bundled stdio bridge

• any MCP client that can talk HTTP or stdio

• direct REST callers when you want curl-first control

Why Auto Browser

• MCP-native from day one. The browser surface is already packaged as an MCP server instead of bolted on after the fact.

• Human takeover when the web gets brittle. noVNC keeps the same live session available when a person needs to step in.

• Login once, reuse later. Save named auth profiles and reopen fresh sessions that are already signed in.

• Local-first by default. Run the full stack on your own box with Docker Compose, or use Codespaces for a quick hosted demo.

• Safety rails built in. Approvals, operator identity, PII scrubbing, Witness receipts, and policy presets are all part of the product surface.

• Evidence you can hand to someone else. Witness receipt chains are Ed25519-signed, and an exported bundle verifies with scripts/verify_witness_bundle.py — which imports nothing from this project, so a recipient need not run or trust this controller to check it. The signing key sits beside the receipts, so for evidence that survives a compromised box, pin a head hash you recorded earlier or ship the receipts off the box (details).

• We audit ourselves in public. docs/audits/2026-08-execution-audit.md documents an adversarial audit of this repo that found safety controls which reported success while doing nothing, with reproductions, the fixes, and the gates that close the class.

• Governed skill induction. Verified browser traces can become staged skill candidates with provenance that is signed when a mesh identity is configured — and checked on read, not just produced — plus verifier adapters and review-only graduation — agents that prove they can repeat themselves correctly, not just act once.

Related MCP server: Browser-Use MCP Server

Good Fits

• internal dashboards and admin tools

• operator-assisted QA and browser debugging

• login-once, reuse-later account workflows

• brittle sites where a human may need to recover the flow

• MCP-powered agent workflows that need a real browser, not just HTML fetches

Not the Goal

• CAPTCHA solving

• unauthorized scraping or account automation

• deceptive identity shaping or bypass tooling

What You Get

Quickstart

git clone https://github.com/LvcidPsyche/auto-browser.git
cd auto-browser
docker compose up --build

That is enough for local development with the default settings. On first start the controller generates an API token and saves it to data/api-token in this folder (git ignores it), and every API and MCP request must send it as Authorization: Bearer <token>. The dashboard asks for it once, the examples below read it with $(cat data/api-token), and the MCP clients take it as shown under MCP Clients. API_ALLOW_TOKENLESS_LOOPBACK=true runs the controller without a token, as it ran before, but then any container on the compose network can drive the browser.

Optional:

cp .env.example .env
make doctor

Run make doctor from a normal terminal with local Docker access and permission to open localhost sockets.

Open:

• API docs: http://127.0.0.1:8000/docs (Authorize with the token to try requests). The docs need no token on a loopback deployment (an exposed one answers 401), and Swagger UI loads its assets from cdn.jsdelivr.net.

• Operator dashboard: http://127.0.0.1:8000/dashboard, which includes the queue of actions waiting on your approval; paste the token when it asks, or open http://127.0.0.1:8000/dashboard#token=<token>

• Visual takeover: http://127.0.0.1:6080/vnc.html?autoconnect=true&resize=scale

All published ports bind to 127.0.0.1 by default.

Prebuilt images

Every release is also published to GHCR, so you can skip the build:

docker compose -f docker-compose.yml -f docker-compose.images.yml up -d

AUTO_BROWSER_IMAGE_TAG picks the version: latest (the default), a release such as 1.10.1, or edge for the current main. The images are linux/amd64, and each carries an SBOM and a build provenance attestation (gh attestation verify oci://ghcr.io/lvcidpsyche/auto-browser-controller:latest -R LvcidPsyche/auto-browser). The controller image leaves out the agent CLIs; for *_AUTH_MODE=cli, build from source with INSTALL_AGENT_CLIS=true.

With isolated sessions, name docker-compose.images.yml last, so they start from the published browser image too:

docker compose -f docker-compose.yml -f docker-compose.isolation.yml -f docker-compose.images.yml up -d

In the other order, the isolation overlay's default wins, and isolated sessions run a local auto-browser-browser-node:latest: a stale build, or no image at all.

Try It in Codespaces

Codespaces provisions the stack automatically. The dashboard and noVNC tabs are usually ready in about 90 seconds.

First Useful Demo

The highest-signal flow in this repo is:

1. create a session

2. log in manually if the site needs a human

3. save the session as a named auth profile

4. open a new session from that auth profile

5. continue work without reauthing

Start here:

• examples/login-and-save-profile.md

• examples/README.md

Minimal session creation, from the auto-browser folder (AUTH carries the token; if you set API_BEARER_TOKEN, use $API_BEARER_TOKEN in place of $(cat data/api-token)):

AUTH=(-H "Authorization: Bearer $(cat data/api-token)")
curl -s "${AUTH[@]}" http://127.0.0.1:8000/sessions \
  -X POST \
  -H 'content-type: application/json' \
  -d '{"name":"demo","start_url":"https://example.com"}' | jq

Minimal observation:

curl -s "${AUTH[@]}" http://127.0.0.1:8000/sessions/<session-id>/observe | jq

Recent Changes

1.10.1

• Rebuilt Docker images with updated dependencies (Pillow 13.0.0). Image tags now publish only from a main commit whose CI run succeeded.

1.10.0

• Operators can no longer reach each other's work. With named credentials (API_BEARER_TOKENS), saved auth state, exported archives, harness runs, skill candidates, memory profiles, uploads, MCP resource lists and controller-wide approvals belong to the operator who made them.

• ALLOWED_HOSTS holds wherever a page ends up: after a redirect, a tab switch, a cookie written through browser.set_cookies, or a page that navigates itself or opens a popup (a popup's first page is held from 1.11.2). A page that gets off the list on its own is sent back to about:blank once it has navigated, not before its request goes out.

• Witness verification rejects a chain with its signatures stripped, and the top-level valid now covers signatures. Read the upgrade notes in the changelog before upgrading.

• Fixes for the Playwright export, stealth mode, REST error statuses, text-only agent steps, private-key scrubbing and the dashboard's error reporting, plus an opt-in FORCE_GOVERNED_APPROVALS.

1.9.1

• Prebuilt Docker images on GHCR (thanks @FRFlo): run a release without building it. See Prebuilt images.

1.9.0

• Operators no longer share sessions. With named credentials (API_BEARER_TOKENS), a session belongs to the operator who created it, as auth profiles already did.

• Every setting in .env reaches the controller. Compose used to forward only the keys it listed, so settings like API_BEARER_TOKENS and PII_SCRUB_* were ignored. Check your .env before upgrading.

• Governed approvals run, a page stuck in a script loop can no longer hang the controller, and the Claude and OpenAI providers work on current models.

• Security fixes for approvals, navigation, TOTP autofill, PII scrubbing and witness bundle verification (GHSA-37hm-f6gf-q7vx).

1.8.1

• Sturdier sessions and stores. Actions that ran are no longer reported as failed when the page navigates right after, two requests carrying one approval no longer both run its action, MAX_SESSIONS holds under concurrent creates, and sessions, tabs, cron jobs and the JSON stores no longer leak or lose writes.

• The stdio MCP bridge survives a controller restart and reports auth and rate-limit errors instead of hanging. The SDK, bridge and LangChain adapters can send an operator id for controllers with REQUIRE_OPERATOR_ID=true.

• Faster actions and observations, fewer browser round trips per observation, and a controller image without the test tooling.

• Security fixes for the noVNC socket, TOTP autofill, auth-profile paths, approvals and witness receipts. create_session with totp_secret now needs totp_hosts or a start_url.

See CHANGELOG.md for the full release history.

MCP Clients

Auto Browser exposes:

• an HTTP MCP endpoint at http://127.0.0.1:8000/mcp

• convenience endpoints at http://127.0.0.1:8000/mcp/tools and http://127.0.0.1:8000/mcp/tools/call

• a stdio bridge: uvx auto-browser-mcp from PyPI, or scripts/mcp_stdio_bridge.py in a repo checkout

Clients that speak MCP over HTTP connect directly, sending the controller's token. With Claude Code or Codex, from the auto-browser folder (--scope user makes the server available in every project, not only this folder):

claude mcp add --transport http --scope user auto-browser http://127.0.0.1:8000/mcp \
  --header "Authorization: Bearer $(cat data/api-token)"

# Codex reads the token from the variable you name: export it wherever Codex runs.
export AUTO_BROWSER_TOKEN="$(cat data/api-token)"
codex mcp add auto-browser --url http://127.0.0.1:8000/mcp --bearer-token-env-var AUTO_BROWSER_TOKEN

Google Antigravity takes the server in its raw MCP config (MCP Servers, then Manage MCP Servers, then View raw config), or in ~/.gemini/config/mcp_config.json. It reads serverUrl, not url, and the token goes in headers:

{
  "mcpServers": {
    "auto-browser": {
      "serverUrl": "http://127.0.0.1:8000/mcp",
      "headers": { "Authorization": "Bearer <contents of data/api-token>" }
    }
  }
}

Cursor, VS Code, Claude Desktop through the stdio bridge (which reads the file named by AUTO_BROWSER_TOKEN_FILE / --token-file; the repo wrapper, make stdio-bridge or scripts/mcp_stdio_bridge.py, defaults that to its checkout's data/api-token for a loopback URL), and pairing Auto Browser with a web-search MCP server are covered in docs/mcp-clients.md.

The default MCP tool profile is curated: the 20 tools a browsing agent needs (sessions, observe and screenshot, execute_action, page and download reading, tabs, auth profiles, human takeover). Every listed tool costs the model context on every request, so the diagnostics, audit, harness, and admin tools are in the full profile. To expose everything, set:

MCP_TOOL_PROFILE=full

MCP_TOOL_PROFILE=lean serves just the 7-tool browsing loop (observe, screenshot, execute_action, get_html, find_elements, human takeover, close), and MCP_TOOL_PACKS adds or removes named packs on top of any profile, e.g. MCP_TOOL_PACKS=tabs,downloads or -scripting. The pack table is in docs/mcp-clients.md.

Raw tool-call example:

curl -s "${AUTH[@]}" http://127.0.0.1:8000/mcp/tools/call \
  -X POST \
  -H 'content-type: application/json' \
  -d '{
    "name":"browser.create_session",
    "arguments":{
      "name":"demo",
      "start_url":"https://example.com"
    }
  }' | jq

Client setup guides:

• docs/mcp-clients.md

• examples/claude-desktop-setup.md

• examples/cursor-mcp-setup.md

• examples/claude_desktop_config.json

For resource listing, resource reads, and subscription-style update examples, see docs/mcp-clients.md#resources-and-subscriptions.

Convergence Harness

Auto Browser ships a Stage 0 convergence harness for Agent Skill Induction. It runs a structured task contract, records tamper-checked traces, verifies completion, and writes a staged skill candidate carrying provenance. With a mesh identity configured that provenance is signed, and the registry verifies the signature before serving a candidate — a candidate that fails the check, or that was dropped into the staging directory unsigned, is refused. Candidates induced from a mock run are marked simulated so they cannot pass as converged. Generated skills are staged only — promotion stays explicit and reviewed.

The harness tools — convergence runs, run status and traces, drift checks, candidate management, and graduation — are in the full MCP tool profile (MCP_TOOL_PROFILE=full), or can be invoked directly over REST.

Start with docs/convergence-harness.md. A deterministic local smoke is:

python -m controller.harness.run --contract evals/contracts/example_read.json --mock-final-url https://example.com --mock-final-text "Example Domain"

For MCP clients, set MCP_TOOL_PROFILE=full (or MCP_TOOL_PACKS=harness on top of curated) to expose the harness.* tools.

Security and Compliance

For a real private deployment, set at least:

APP_ENV=production
API_BIND_SCOPE=exposed
API_BEARER_TOKEN=<strong-random-secret>
REQUIRE_OPERATOR_ID=true
AUTH_STATE_ENCRYPTION_KEY=<44-char-fernet-key>
REQUIRE_AUTH_STATE_ENCRYPTION=true
SHARE_TOKEN_SECRET=<strong-random-secret>
CONTROLLER_ALLOWED_HOSTS=<controller hostname>
ALLOWED_HOSTS=<sites the browser may visit>
REQUEST_RATE_LIMIT_ENABLED=true
METRICS_ENABLED=true
STEALTH_ENABLED=false

With APP_ENV=production the controller refuses to start while any of these is missing, and the startup log names the missing ones.

By default every session shares one Chromium process and one noVNC desktop (SESSION_ISOLATION_MODE=shared_browser_node). Cookies and storage stay separate per session, but a human taking over one session can see the others' windows. When sessions belong to different people, accounts or trust domains, start with make up-isolation to give each session its own browser container, takeover surface and Docker network (docker_ephemeral), so one session's browser cannot reach another's Playwright server. docs/session-isolation-audit.md has the details.

ALLOWED_HOSTS is held before the request goes out (PRE_REQUEST_BLOCKING=true, the default):

• Stopped before it is sent: a top-level navigation's request to a host off ALLOWED_HOSTS, including every redirect hop and a popup's first request.

• Still caught after the fact by NAV-07 (the guard that sends a refused page back to about:blank once it commits):

  • speculation-rules prefetch in attach modes, and in a launched browser where it could not be turned off (launched browsers turn it off when the controller acquires them; /readiness warns when that failed);

  • a navigation a service worker serves (one a same-site refused iframe registered) for a popup's first navigation, and for any the popup makes before its service-worker bypass is set (a background task once the controller sees the popup), in every mode;

  • blob: navigations a refused iframe makes;

  • bfcache restores in attach modes;

  • requests held when the gate detaches, and any navigation a page starts itself before the replacement gate attaches. They go out. The controller attaches the replacement before it closes that browser's sessions, so the window is one attach, and it refuses their actions meanwhile.

• Any tab, rarely: a navigation to a refused host fails closed in any tab of a gated browser, the operator's own included in attach modes, when the gate cannot classify its frame (its lookup failed).

• Not covered: iframes and subresources of other hosts.

In an attach mode (CDP_CONNECT_URL, browser.cdp_attach) the gate stops navigations of the controller's sessions only (apart from a frame it cannot classify, above), but every navigation in that Chrome, the operator's own tabs included, waits on its reply: about 13 ms each, the median measured on Chromium 151. Start that Chrome with --disable-features=BlockOriginHeaderModificationOnRedirect,AvoidCorsURLLoaderRestartOnRedirect and set CDP_BLOCK_ORIGIN_ON_REDIRECT_DISABLED=true. Without it, cross-origin 307/308 POST redirects can fail, and on Chrome 155 and later credentialed cross-origin requests that need a CORS preflight fail after a redirect, in the operator's own tabs too, while the controller is attached. Details: docs/deployment.md.

COMPLIANCE_TEMPLATE can apply a preconfigured posture at startup:

Both presets require upload approvals and enable Witness receipts. Startup writes the applied policy to /data/compliance-manifest.json. The legacy names (HIPAA, SOC2, GDPR, PCI-DSS) still work as deprecated aliases and emit a warning at startup.

Example:

COMPLIANCE_TEMPLATE=strict docker compose up

For deployment details, hosted Witness notes, CLI auth modes, and reverse-SSH guidance, see:

• docs/deployment.md

• docs/production-hardening.md

• docs/session-isolation-audit.md: what each isolation mode does and does not separate

Architecture at a Glance

flowchart LR
    User[Human operator] -->|watch / takeover| noVNC[noVNC]
    LLM[Any model: OpenAI / Claude / Gemini / OpenRouter / Grok / DeepSeek / MiniMax / local] -->|shared tools| Controller[Controller API]
    Controller -->|Playwright protocol| Browser[Browser node]
    noVNC --> Browser
    Browser --> Artifacts[(screenshots / traces / auth state)]
    Controller --> Artifacts
    Controller --> Policy[Allowlist + approval gates]

Model providers

First-class adapters for OpenAI, Claude, and Gemini (API or CLI). Beyond those, a single generic OpenAI-compatible adapter drives any model reachable over an OpenAI /chat/completions endpoint — set an API key to enable it:

Vision (screenshots) is used for every provider except text-only ones. See .env.example for the *_API_KEY / *_BASE_URL / *_MODEL settings.

Core components:

• browser-node/ runs Chromium, Xvfb, x11vnc, and noVNC

• controller/ exposes the FastAPI controller, MCP transport, policy rails, and orchestration endpoints

• data/ holds runtime artifacts, auth state, approvals, audit logs, and optional CLI caches

• scripts/ contains local helpers for doctor, smoke tests, bridges, and release checks

Repo Guide

Common Commands

Documentation Map

Contributing

If you want to help, start with:

• CONTRIBUTING.md

• docs/good-first-issues.md

• CODE_OF_CONDUCT.md

If Auto Browser is useful, a star helps other people find it. Sponsorship and tip options live in TIPS.md.