Browser MCP

One shared browser profile with separate tab ownership for OpenCode sessions and Codex conversations. The plugin wraps chrome-devtools-mcp and routes each conversation through its own browser tabs while sharing cookies, logins, local storage, cache, and extensions.

Package 2.x targets OpenCode v2 (2.0.27 or newer). Package 1.x retains the OpenCode v1 adapter (1.18 or newer, below 2). Codex uses the same Rust MCP server and manifests in both releases. The native v2 adapter uses @opencode/plugin; local directories also expose a physical server entrypoint.

Install @mirsella/browser-mcp for OpenCode v2, or @mirsella/browser-mcp@opencode1 for OpenCode v1. Existing Nix-managed and local-directory installations can keep their configured paths.

Requirements

• OpenCode 2.0.27 or newer, or Codex with plugin and hook support

• Linux x64

• Bun

• A local Chromium-family browser

• A graphical session, unless OPENCODE_BROWSER_HEADLESS=true

Related MCP server: browser-mcp-cdp

OpenCode setup

Add the plugin to opencode.json. It enables the browser server and makes the browser automation skill available by default:

{
  "$schema": "https://opencode.ai/v2/config.json",
  "plugins": ["@mirsella/browser-mcp"]
}

Restart OpenCode after changing its configuration. That is the complete setup. You do not need a separate installation, broker command, or background service.

The first adapter starts the packaged native broker automatically. Concurrent Codex and OpenCode instances converge on the same private socket and browser profile. The broker remains available while at least one adapter is connected, then closes the browser and exits shortly after the final adapter disconnects.

The CDP port defaults to 9222. If that port is occupied, the broker logs the collision and selects a free loopback port without adopting or stopping the other listener. Adapters keep the same configuration and share the selected endpoint. A replacement broker reuses that port only when the exact persisted browser generation is still running.

OpenCode loads the JavaScript ./server package export. Codex uses the plugin and MCP manifests plus the session hooks. Both integrations launch the same packaged Rust executable.

Timeouts and diagnostics

Each browser action has one deadline for catalog discovery, browser startup, session locking, execution, and reconciliation. OPENCODE_BROWSER_OPERATION_SECONDS sets that deadline and defaults to 540 seconds. Catalog discovery and compact help have a separate 70-second limit. A failed or panicking catalog discovery can be retried without restarting the broker.

After an interrupted action, the next call replaces that session's worker and reconciles its existing owned tabs. Actions are never replayed automatically. Dispatched tab creation finishes registering ownership or closing that exact target after caller cancellation. Inspect the page before retrying a write because it may already have taken effect.

The broker writes diagnostics to $XDG_STATE_HOME/browser-mcp/broker-<profile-hash>.log, or ~/.local/state/browser-mcp/ when XDG_STATE_HOME is unset. Its startup message prints the exact path. Logs are readable only by your user and also go to stderr.

Request tracing records the tool name, session ID, current stage, and elapsed time. Chrome DevTools worker stderr is drained into that private broker log, so replacement workers survive a disconnected host's stderr. Request tracing does not record tool arguments. If a call stalls, look for its last stage and whether the broker recorded a deadline, cancellation, catalog failure, or worker replacement. Browser stderr is captured in browser-<profile-hash>.log in the same directory. Its separate writer survives broker replacement and exits when Chromium closes the pipe. Each broker or browser log retains an 8 MiB file and one rotated .log.1 backup. A browser that exits during startup reports that exit immediately.

Codex setup

The release tarball contains the executable, both integrations, and a local Codex marketplace. Download it from GitHub Releases, extract it to a permanent directory, then register that directory:

codex plugin marketplace add /path/to/browser-mcp
codex plugin add browser@browser-mcp

The npm package is now @mirsella/browser-mcp; it replaces @mirsella/opencode-browser-mcp. To use a GitHub release in OpenCode, set the plugin entry to the file:///absolute/path of its unpacked directory instead of the npm name.

For a source checkout, build and stage the binary first (bun run build:native and bun run stage:native). Use a packaged directory for installation so compiler outputs are not copied into Codex's plugin cache.

The installed plugin is enabled by default. Open /hooks in Codex to review and trust its two hooks, then restart Codex. They inject the conversation ID before browser calls and release owned tabs at SessionEnd. Browser calls without an injected identity fail closed. The plugin exposes compact help and dispatch tools and loads the same browser automation skill as OpenCode.

Codex and OpenCode use separate ownership namespaces while sharing the existing broker, profile, cookies and logins. Codex's hook API identifies subagents with their parent conversation ID, so subagents share that conversation's tabs. Use separate pages when they work concurrently.

Codex runs SessionEnd on normal shutdown, on archive/delete of an open conversation, or after 30 minutes idle with no client connected. Its hook has a three-second deadline: the broker acknowledges cleanup promptly, then waits for active calls before closing only that conversation's owned tabs. Cleanup does not start a broker when none is running. Crashes still use normal broker shutdown and idle cleanup.

The Codex MCP entry uses a 540-second tool timeout. Add --direct-tools to its arguments when configuring per-action Codex tool approval rules; compact dispatch has one approval rule for all actions.

Plugin behavior

The plugin:

• registers and enables the packaged browser MCP adapter;

• injects the active OpenCode session ID after model argument generation;

• exposes compact browser tools unless existing permissions require direct tools;

• releases owned tabs when an OpenCode session is deleted;

• registers the bundled browser automation skill when the server is enabled.

The adapter preserves OpenCode's MCP request metadata and routes calls by _meta["ai.opencode/sessionID"], including calls that bypass the plugin tool hook. Legacy plugin routing and Codex hooks remain supported. Conflicting or missing identities fail closed; unidentified callers never share a browser session. The broker still binds each session to its original workspace and keeps tab ownership separate.

OpenCode sees two compact tools by default:

• browser_dispatch executes one browser action.

• browser_help lists actions or loads one action schema.

If the plugin detects per-action browser_* permission or tool rules, it uses direct tools so those policies continue to work.

Settings:

• enabled (default true): start the server and load the skill. Set it to false to disable both. An explicit mcp.servers.browser entry takes precedence.

• compactTools (default true): keep browser_dispatch and browser_help instead of the direct tools. Set it to false to always use direct tools:

{
  "$schema": "https://opencode.ai/v2/config.json",
  "plugins": [
    { "package": "@mirsella/browser-mcp", "options": { "compactTools": false } }
  ]
}

Agent workflow and context

The initial MCP catalog remains two tools. Help returns one short signature and argument schema. Structured upstream data is validated internally, then the agent receives one readable copy plus any images. Text is limited to 32 KiB with an explicit truncation notice. Network and console lists default to 50 entries.

• list_pages exposes stable session-local handles such as p1. select_page changes current; optional page on scoped actions targets another owned page without changing current. Closed handles fail rather than selecting another tab.

• take_snapshot adds find, within, interactive, diff, and maxLines (default 300, maximum 2000). Search includes ancestors. Diff requires compatible document/worker/verbosity state; an unavailable baseline is explicitly reported and replaced with a bounded full view. Baselines use at most 2 MiB per session.

• Inline snapshot UIDs include a worker namespace. Old-worker references fail before execution. Saved snapshot files retain upstream UIDs and are diagnostic exports; obtain actionable references from an inline snapshot.

• Compact-only batch runs up to 8 known actions under one session lock and deadline. All arguments, explicit pages and existing UIDs are checked first. Only the final step may return a snapshot. It stops on the first error and reports completed/not-run counts. Actions are never replayed automatically.

• handoff with mode: "start" focuses an exact headed page for user work and returns immediately. End the turn, request the user's work, then finish only after confirmation. Automation on that page pauses and normal idle cleanup is delayed for at most 24 hours. Explicit session release still closes owned tabs.

• status observes existing state without starting a browser or worker. browser-mcp doctor checks local configuration without starting services.

Ordinary action errors retain a healthy, reconciled worker and its element references. Transport failures, interrupted calls and failed reconciliation still replace the worker. New pages default to background; per-session window placement may briefly activate an existing anchor window.

Recording is explicitly enabled with --recording or OPENCODE_BROWSER_RECORDING=true. It requires ffmpeg; --ffmpeg or OPENCODE_BROWSER_FFMPEG selects its executable. Only then do screencast_start and screencast_stop appear. Artifacts are assigned under the conversation workspace's .scratch/browser-mcp/. Stop recording before cleanup. Worker loss can leave an incomplete file. Recordings use WebM. The encoder launcher rounds padding to even dimensions to fix the pinned recorder's odd-viewport failure; recording operations stop waiting after 30 seconds and report an incomplete recording. There is no automatic or rolling capture.

Optional --secrets-file / OPENCODE_BROWSER_SECRETS_FILE enables fill_secret. The existing file must be a same-user, nonsymlink, regular file with mode 0600, at most 64 KiB. No store is created. Its format is:

{"account_password":{"value":"example-placeholder","origins":["https://example.com"]}}

Each name has a nonempty single-line value up to 4096 bytes and exact HTTPS origins. fill_secret takes a name, an editable inline-snapshot UID and optional page handle. It checks the top-level origin in the same synchronous function as insertion and returns no value. Treat recordings and exported artifacts as sensitive; see SECURITY.md.

HAR export is deferred: the current worker does not expose the timing data needed for an accurate HAR. Use paginated network lists and get_network_request with requestFilePath / responseFilePath for focused diagnostics instead.

Architecture

flowchart LR
    O1[OpenCode instance A] --> A1[packaged stdio adapter]
    O2[OpenCode instance B] --> A2[packaged stdio adapter]
    A1 -->|private Unix socket| R[auto-managed broker]
    A2 -->|private Unix socket| R
    R --> T[CDP target tracker]
    R --> W[session worker pool]
    W --> C[chrome-devtools-mcp]
    T --> B[managed browser]
    C --> B
    B --> P[shared locked profile]

The adapter starts the broker on demand but does not own it. This lets adapters from separate OpenCode processes share one browser without tying its lifetime to whichever process connected first. Socket and profile locks make simultaneous startup race-safe and prevent two browsers from opening the same profile.

Each routed session leases its own chrome-devtools-mcp worker. Stable browser-level CDP target IDs remain the ownership authority; worker-local page numbers are removed from model-visible schemas and responses.

Ownership guarantees

• A tab has at most one owner.

• Routed sessions execute concurrently through separate workers.

• Popups inherit the owner of their opener tab.

• A session cannot select, inspect, or close another session's tabs through the broker.

• Manual and otherwise unowned tabs are not claimed during cleanup.

• Ownership conflicts, stale browser generations, and malformed responses fail closed.

These are routing guarantees, not an operating-system security boundary. Any process running as the same Unix user can connect directly to the browser's loopback CDP port.

Browser and profile

The plugin detects Chromium, Google Chrome, Helium, Brave, and Microsoft Edge. Chromium is preferred by default; OPENCODE_BROWSER_EXECUTABLE overrides automatic detection. Google Chrome 136 and newer ignore remote-debugging flags for its default data directory, so Chrome uses a dedicated directory under ~/.config/opencode-browser-mcp by default. Other detected browsers retain their existing profile defaults.

The broker starts Chromium lazily on the first browser action. For a headed browser it passes only a fixed allowlist of display and desktop variables. The current process environment stays authoritative; missing values are recovered from the current user session when available.

Configuration variables inherited from either host (the existing names are shared by both integrations):

Lifecycle

Calls interrupted after dispatch can have taken effect. They return an error and are never replayed automatically. Inspect the browser state before retrying an action. Explicitly disabled or disconnected MCP entries stay disabled.

Troubleshooting

• A headed startup error mentioning DISPLAY or WAYLAND_DISPLAY means no graphical session could be found. Set OPENCODE_BROWSER_HEADLESS=true when running without one.

• A profile or socket lease error means another live package process owns it or a stale process is still running. Do not bypass the locks.

• If startup reports that the profile is already open in another local process, that browser may have been opened without remote debugging. Save your work, close it gracefully, and retry the MCP action. MCP will start the same profile with remote debugging enabled. To leave that browser running, use a separate OPENCODE_BROWSER_PROFILE or --profile instead. Never delete Chromium's SingletonLock or kill the existing browser automatically.

• A process on the configured CDP port must belong to the exact managed browser generation. Stop it or choose another port.

• If downstream startup fails, verify that bunx is available to OpenCode.

• The plugin refuses to replace an unrelated mcp.servers.browser entry.

Security

All sessions deliberately share cookies, authentication, storage, extensions, permissions, history, and downloads. Browser content remains untrusted and may contain prompt injection. The private same-UID socket prevents access by other local users, but it does not isolate processes already running as your Unix user.

File exports can use the conversation's workspace and /tmp/opencode. Chrome DevTools also permits its OS temporary directory. /tmp/opencode is a shared scratch directory for small, temporary snapshots, screenshots, and network responses. Clean up exported files when finished; store larger or lasting artifacts in the workspace.

Read SECURITY.md before using a profile with sensitive data.

Development

cargo fmt --check
cargo test --all-targets
cargo clippy --all-targets -- -D warnings
bun test tests/plugin.test.ts
cargo test --release --test e2e -- --ignored --nocapture --test-threads=1

The headed E2E starts adapters against temporary profiles, sockets, and CDP ports. It checks automatic startup, startup races, cookie sharing, tab isolation, concurrent calls, active-call cleanup, reconnect grace, compact dispatch, both window layouts, and automatic browser shutdown.

See CONTRIBUTING.md for the development and release workflow.

Releases

CI builds and tests the distributable tarball on every push. Run the release workflow from main with the full tested source commit: main for 2.x, or release/opencode-v1 for 1.x. The workflow verifies that commit and downloads its checksummed CI artifact instead of rebuilding it. Dispatch with npm_mode=publish to publish through npm's OIDC trusted publisher, create the GitHub release, and update the Codex marketplace branch for latest. Use npm_mode=dry-run to verify without writing. The channels are latest for 2.x and opencode1 for 1.x.

License

MIT