chromium-bridge
Bridges Model Context Protocol clients to your actual Chromium browser, enabling control of tabs, reading authenticated pages, cookies, and storage, and running JavaScript on approved sites. Emphasizes security with per-site approval, high-risk action confirmations (including Touch ID on macOS), and an attested native-messaging bridge.
README
chromium-bridge
Let any MCP client (Claude Code, Claude Desktop, Codex, or anything that
speaks the Model Context Protocol) drive your real Chromium browser: your
tabs, your logged-in sessions, your cookies, through a browser extension and
a native-messaging host. No second browser, no CDP debug port, no
--remote-debugging flag.
Because it operates the browser you are already signed into, an agent can do
things a fresh headless browser cannot: read a page behind your auth, click
through an app you are logged into, pull a token your framework stashed in
localStorage. That power is also the risk. Read the security section before
you install.
Translations: Simplified Chinese, Traditional Chinese.
Security first
chromium-bridge drives a real, authenticated browser. It can read page
content, cookies (including httpOnly), and web storage, and can run
JavaScript in your pages. The guardrails:
- Approve every site. A new origin triggers a prompt; nothing runs on a site you have not approved.
- Confirm high-risk actions. Submit clicks, key presses, tab close, file
uploads, and every
page_evalconfirm on an extension-owned window the page cannot see or click. On a Mac enrolled via Touch ID,page_evalandpage_uploadapproval is a Secure Enclave user-presence check (Touch ID or the login password) that no page or program can forge (ADR-0031). These gates are on by default; each is a documented setting, and relaxing one is an explicit, informed choice (SECURITY.md). - Read-only credentials. Cookies and storage can be read (always masked:
JWTs, long hex, long digit runs), never written. There is no
cookie_setorstorage_setby design. - Authenticated, attested bridge. On macOS and Linux the bridge between the host processes is a private Unix-domain socket (no listening port). Every connection must pass a kernel peer-UID check, kernel-attested executable identity, and an HMAC challenge over a per-run secret. MCP clients themselves are admitted against a trusted-client allowlist keyed on attested code identity, and any side can revoke trust at any time (ADR-0024, ADR-0025).
- A global kill switch. One action from the CLI, the extension, or the app halts everything until you release it with proof of presence (ADR-0030). Every security decision lands in an on-disk audit trail.
Platform honesty: the strong bridge guarantees (portless socket, peer-UID check, attestation) exist on macOS and Linux only. On Windows the bridge is a loopback TCP socket gated only by the HMAC secret, and the server warns about this at startup. Windows support is best-effort. Details in SECURITY.md.
Full details: SECURITY.md, threat model, trust boundaries, per-tool risk matrix.
Quickstart with the app (macOS)
The Chromium Bridge desktop app is the primary install path. It bundles the signed host binary and the extension, and the only command in this path is the one that registers the server with your MCP client.
App downloads are not published yet: releases carry the CLI archive and extension zip, and the release pipeline's desktop job stays dormant until its signing secrets are configured and the publish hold is lifted (see docs/release.md). Until then, build the app from a source checkout:
moon run dmg-appproduces a signed disk image,moon run install-appputs the built app in /Applications, andmoon run run-appbuilds and launches it in place. A build signed with the free development certificate runs only on Macs its provisioning profile lists, and public distribution also needs a paid Developer ID for notarization; both remain open. Or use the CLI path below.
-
Install the app. Get
Chromium Bridge.app(see the note above) and open it. On first launch it registers the native-messaging host with every Chromium browser it detects (Chrome, Brave, Edge, ...) and shows what it wrote. -
Load the extension. On the app's Setup page, click "Reveal folder" to open the bundled extension, then in your browser open
chrome://extensions, enable Developer mode, click "Load unpacked", and select that folder. Restart the browser so it picks up the registration. -
Pair with Touch ID. On the app's Pairing page, click Pair (a Touch ID prompt appears), then approve the key fingerprint on the extension's options page. On macOS the extension requires this enrollment by default and refuses to act until the fingerprints match (ADR-0021).
-
Connect your MCP client. On the Setup page, click Install to put the
chromium-bridgecommand at~/.local/bin. Then register it with your client. For Claude Code:claude mcp add chromium-bridge -- "$HOME/.local/bin/chromium-bridge"Other clients take the same binary as an
mcpServersentry; see Connect your MCP client.
Ask your client to "list my browser tabs". The first time you target a new site, click the Chromium Bridge toolbar icon and approve it.
The app is also the ongoing control panel: pairing, trusted clients, the kill switch, and the audit trail, with the dangerous acts gated by Touch ID on an enrolled Mac. See docs/desktop-app.md.
Quickstart with the CLI (macOS, Linux, Windows)
The CLI is co-equal: everything the app does, from a terminal, with no dependency beyond the binary itself. It is the natural path on Linux and Windows, on headless machines, and in CI.
-
Download the archive for your platform from the latest release and extract it. Optionally verify it first; on macOS/Linux (Windows archives are
.zip, checked with your own sha256 tooling):shasum -a 256 -c chromium-bridge-<tag>-<platform>-<arch>.tar.gz.sha256 gh attestation verify chromium-bridge-<tag>-<platform>-<arch>.tar.gz --repo Vivswan/chromium-bridgeThe full verification story is in SECURITY.md.
-
Register the extracted binary with your browsers. Registration is idempotent, so the same command is the fresh install, the repair, and the re-register after moving the binary:
./chromium-bridge doctor --fix # every detected browser ./chromium-bridge doctor --fix --browser chrome,braveKeep the binary at a stable path (it is registered in place). On Linux,
~/.local/lib/chromium-bridge/is a good home.chromium-bridge uninstallreverses exactly what was registered. -
Load the extension: the archive's
extension/dist/directory viachrome://extensions, Developer mode, "Load unpacked". Restart the browser. -
On macOS, pair:
chromium-bridge pair(Touch ID), then approve the fingerprint on the extension's options page; the extension requires enrollment there by default. Linux and Windows skip this step. -
Connect your MCP client to the extracted binary (absolute path), as below.
Building from source instead: cargo build --release, then run the same
doctor --fix from target/release/chromium-bridge
(see docs/development.md).
The full CLI (doctor, pairing, revocation, kill switch, audit) is documented in docs/cli.md.
Connect your MCP client
Point your client at the installed binary. Run with no arguments, it speaks
MCP over stdio. Use an absolute path; most clients do not expand ~.
Claude Code:
claude mcp add chromium-bridge -- "$HOME/.local/bin/chromium-bridge"
Claude Desktop and other mcpServers JSON clients:
{
"mcpServers": {
"chromium-bridge": {
"command": "/ABSOLUTE/PATH/TO/chromium-bridge",
"args": []
}
}
}
Codex (~/.codex/config.toml):
[mcp_servers.chromium-bridge]
command = "/absolute/path/to/chromium-bridge"
args = []
Several clients can be connected at once: the first server instance becomes a broker and later instances attach to it, each one attested and individually revocable (ADR-0024).
On WSL: if your everyday browser is Windows Chrome, install on Windows and
point the WSL client at the .exe via /mnt/c; do not install a Linux host.
If Chrome runs under WSLg, install natively in Linux. See the
WSL guide.
What you can do: 26 tools
Grouped from the single source of truth, the Rust tool catalogue
(src/packages/core/src/tools/catalogue.rs).
Full blast-radius detail per tool is in the
tool risk matrix.
Browsers
| Tool | Does | Risk |
|---|---|---|
list_browsers |
List the browsers connected to the bridge (label + open-tab count) | low |
Several browsers can be connected at once (on macOS/Linux each gets its own
native host and label, e.g. chrome and brave). Every other tool takes an
optional browser argument to pick one; with several connected, an
unaddressed call fails with a clear error rather than guessing which
logged-in browser to act in. See
ADR-0022.
Tabs
| Tool | Does | Risk |
|---|---|---|
tab_list |
List open tabs (id, title, url, active) | low |
tab_focus |
Bring a tab to the foreground | low |
tab_open |
Open a URL in a new tab (host must be allowlisted) | medium |
tab_close |
Close a tab (confirmation window) | high |
Navigate
| Tool | Does | Risk |
|---|---|---|
page_navigate |
Load an http(s) URL in the active tab | medium |
page_back / page_forward |
Step through history | low |
page_reload |
Reload the active tab | low |
Inspect a page
| Tool | Does | Risk |
|---|---|---|
page_snapshot |
Accessibility-style tree of interactive elements, each with a stable ref |
low |
page_snapshot_precise |
Authoritative a11y tree via chrome.debugger (shadow DOM / complex ARIA); refs use a p prefix |
medium |
page_text |
Visible page text (passwords and card-like numbers masked) | medium |
page_screenshot |
Visible viewport as a PNG | medium |
console_get |
Recent console output, masked | medium |
Drive a page
| Tool | Does | Risk |
|---|---|---|
page_click |
Click by ref or selector; submit/link clicks require confirmation |
high |
page_fill |
Type into a field (native setter, so React/Vue detect it) | high |
page_press |
Send a key or combo (confirmation) | high |
page_select |
Choose an option in a <select> (confirmation) |
high |
page_hover |
Move the pointer over an element | low |
page_scroll |
Up / down / top / bottom / N pixels | low |
page_wait_for |
Wait for a selector, text, or navigation | low |
page_handle_dialog |
Accept or dismiss a JS dialog (off by default) | high |
Run code and upload (highest risk)
| Tool | Does | Risk |
|---|---|---|
page_eval |
Execute arbitrary JS. Every call confirms, showing the full code; Touch ID on an enrolled Mac. Return value masked by default. Prefer the tools above. | critical |
page_upload |
Attach a named local file to a file input (off by default; every call confirms with the path) | critical |
Read credentials (read-only, always masked)
| Tool | Does | Risk |
|---|---|---|
cookie_get |
Read cookies for the active tab, incl. httpOnly; allowlisted hosts only |
high |
storage_get |
Read the page's localStorage / sessionStorage (same-origin) |
high |
No write tools by design; cookie/storage writes are out of scope (ADR-0010).
How it works
One Rust binary, two modes, joined by an authenticated local socket. A desktop app and the CLI manage the same state.
MCP client A --stdio--> chromium-bridge (broker: first MCP server instance)
MCP client B --stdio--> chromium-bridge ----attach----^ |
(each client attested against the trusted-client | bridge socket
allowlist before it is served) | (Unix-domain socket,
| peer attestation +
| HMAC; TCP on Windows)
v
chromium-bridge --native-host <-- spawned by
| each browser
| chrome.runtime.connectNative
v
Chromium Bridge extension (MV3) --> your page
- MCP server (default mode): launched by your MCP client over stdio.
Speaks JSON-RPC 2.0 (MCP protocol
2026-07-28, stateless, with temporary legacy compatibility for older harnesses). The first instance owns the socket and becomes the broker; later instances attach as relays, so several clients share the browsers concurrently. --native-host: launched by the browser via the host manifest. A thin bridge translating Chrome's native-messaging frames to NDJSON on the socket. Each installed browser launches its own host with its own label, so one broker can address several browsers by name.- Desktop app / CLI: co-equal management surfaces over the same core (registration, pairing, revocation, kill switch, audit). Neither is a trust root; capability-granting acts end in a user-presence gate.
Why two processes? The browser spawns the native host; the MCP client spawns the server. They are not parent and child, so they need an IPC. The native host stays thin so that MV3 service-worker recycling (about every 5 minutes) and host restarts do not lose session state.
Deep dive: docs/architecture.md.
Compatibility
| Supported | |
|---|---|
| macOS | Apple Silicon (arm64) prebuilt; the desktop app and Touch ID gates live here. Intel builds from source. |
| Linux | x64 prebuilt; any Chromium-based browser; CLI management surface. |
| Windows | x64 prebuilt (native, no admin). Bridge security is best-effort; see SECURITY.md. |
| Browser | Any Chromium-based browser, Manifest V3 |
| MCP protocol | 2026-07-28 (ADR-0034) |
| Internal bridge protocol | 1 (BRIDGE_PROTOCOL_VERSION in src/packages/core/src/protocol.rs) |
Known browsers (--browser keys): chrome, chromium, brave, edge,
vivaldi, opera. Every Chromium browser reads the same native-messaging
manifest; only the per-user NativeMessagingHosts location differs, and the
shared resolver in the core knows them all. For a Chromium variant not in the
table, doctor --fix --manifest-dir <dir> targets its directory explicitly
(macOS/Linux; a Windows registry escape hatch is a tracked follow-up).
See docs/compatibility.md and
docs/cli.md.
Configuration
Environment variables read at launch:
| Var | Values | Default | Effect |
|---|---|---|---|
BB_LOG |
error | warn | info | debug |
info |
stderr log / audit threshold |
BB_LOG_FORMAT |
text | json |
text |
Audit-line format; json emits one object per line |
The durable audit trail (chromium-bridge audit) records independently of
these; see docs/cli.md.
Troubleshooting
Run the built-in read-only self-check first:
chromium-bridge doctor # or: chromium-bridge status
It reports whether the server is reachable, the lock-file state, the kill
switch, and each browser's registration state, and doctor --fix repairs
registrations in place. Then check your MCP client's server UI (reconnect
via /mcp in Claude Code) and the extension's service-worker console at
chrome://extensions (look for [bb] logs). Full runbook:
docs/cli.md and docs/operations.md.
Docs map
| Doc | What's in it |
|---|---|
| docs/quickstart.md | Install and first use, app and CLI (also in Simplified and Traditional Chinese) |
| docs/architecture.md | Components, data flow, protocols, security model, key constraints |
| docs/security/ | Threat model, trust boundaries, tool risk matrix, incident response |
| docs/cli.md | The full CLI: doctor/--fix, uninstall, pairing, revocation, kill switch, audit |
| docs/desktop-app.md | The desktop app: what it manages and how to verify it |
| docs/operations.md | The binary modes, logging/audit, the runtime directory, reconnect |
| docs/compatibility.md | Version discipline and the capability/protocol handshake |
| docs/release.md | Release-please releases, prebuilt archives + checksums, SBOM |
| docs/wsl.md | The two WSL modes: Windows Chrome interop and WSLg |
| docs/adr/ | Architecture Decision Records: every "why was this chosen" |
<details> <summary>Testing and project layout</summary>
Independent suites across two languages:
- Protocol layer:
tests/protocol/e2e.py(plusadversarial.pyandchaos.py) drive the real binary over the actual wire protocols. - DOM layer:
tests/browser/dom_test.tsinjects the real content script into an isolated Chrome via CDP and exercises every op against a real DOM. - Smoke:
tests/browser/ext_test.tsboots an isolated Chrome with the built extension. - Real integration (opt-in):
tests/browser/integration_e2e.tswithBB_REAL_E2E=1.
See tests/README.md. Layout: src/apps/host (the Rust
binary), src/apps/extension (MV3, WXT), src/apps/desktop (Tauri app),
src/packages/core (the Rust library and the single source for cross-process
contracts), src/packages/shared (generated TS contracts + validators).
</details>
Project status
Pre-1.0 (Cargo.toml). The protocol layers are covered by end-to-end, adversarial, and chaos tests; the wire parsers are fuzzed. See CHANGELOG.md.
Contributing and governance
CONTRIBUTING.md (workflow), GOVERNANCE.md (how changes get made), SECURITY.md (reporting + review bar), docs/development.md (build/test/release loop).
License
Individual and Small Organization License 1.0.0. Incorporates code from browser-bridge under Apache-2.0; see LICENSE-APACHE and NOTICE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.