Nexus Browser MCP

English | 简体中文

nexus-browser-mcp

A browser-automation MCP server with event-driven, deterministic snapshots.

Built on Playwright. Drives a browser for LLMs through the Accessibility Tree — navigate, click, type, read, fill forms, manage tabs. Key differences from alternatives (e.g. Playwright MCP):

1. Deterministic snapshots: no fixed-interval sleep guessing. A MutationObserver records the last DOM mutation and the browser's own requestAnimationFrame loop decides when the page has been quiet for STABLE_WINDOW_MS (default 800ms) before extracting a snapshot — eliminating "captured mid-animation" races.
2. Built-in governance gates: HITL rules (e.g. clicking "pay/confirm" requires human approval), browser_evaluate disabled by default with unconditional confirmation, JSONL audit log (sensitive-parameter redaction + per-call in/out character metering, so token cost can be reconciled).
3. Multi-task isolation: one MCP connection (session) can host multiple independent task_ids, each with its own BrowserContext (no login-state cross-contamination). Idle tasks are reclaimed by TTL; on next use they're rebuilt and the last page is restored automatically.
4. Death observability + self-healing: if a tab or the whole browser is closed externally or crashes, the next call rebuilds it automatically (a persistent profile keeps your login state) and prepends a [state change] notice telling the agent exactly what was restored and what was lost — no raw Playwright exceptions leak through.
5. Developer observability: every page records console messages, uncaught JS exceptions and network request metadata (method/URL/status/failure reason — never bodies) into capped ring buffers; browser_console / browser_errors / browser_network read them incrementally via a since cursor, so the agent can answer "why did nothing happen" instead of guessing.

Installation

pip install nexus-browser-mcp

or

uvx nexus-browser-mcp

Two executable entry points are installed: nexus-browser-mcp and nexus-browser. A guaranteed fallback: python -m nexus_browser.server.

Requires playwright and its browser binary:

pip install playwright && playwright install chromium

Integrate (any MCP client)

opencode (~/.config/opencode/opencode.json):

{

"mcp": { "browser": { "type": "local", "command": ["uvx", "nexus-browser-mcp"], "enabled": true } } }

Claude Code (.mcp.json, project root):

{

"mcpServers": { "browser": { "type": "stdio", "command": "uvx", "args": ["nexus-browser-mcp"] } } }

Pi Coding Agent: reads standard MCP configuration — project .mcp.json or user-global ~/.config/mcp/mcp.json; stdio is the default transport:

{

"mcpServers": { "browser": { "command": "uvx", "args": ["nexus-browser-mcp"] } } }

See docs/INTEGRATE.md for details (Chinese).

Use your own browser (with login state)

By default, isolated mode launches Playwright's bundled Chromium without your cookies/login state. To use your own browser, pick one:

Option A — load your browser profile directly (recommended, simplest)

Use system Chrome with your everyday user data directory (cookies/login/bookmarks included):

BROWSER_CHANNEL=chrome

BROWSER_USER_DATA_DIR="C:\Users\ \AppData\Local\Google\Chrome\User Data"

Note: while running against your real User Data, the process owns the browser — launching your own Chrome concurrently will conflict. Prefer a copied profile or a dedicated --user-data-dir.

Recommended: a tool-dedicated profile (no conflict with your daily browser)

Use BROWSER_CHANNEL=chrome plus a dedicated user data dir (e.g. C:\Users\ \.nexus-browser\chrome-profile):

BROWSER_CHANNEL=chrome

BROWSER_USER_DATA_DIR="C:\Users\ \.nexus-browser\chrome-profile"

On first use, log in to target sites once in the dedicated Chrome window that pops up when the agent calls a browser tool. Cookies persist in that profile forever after — the agent carries login state while staying fully isolated from your daily browser.

Option B — attach to a running Chrome via CDP

Start chrome --remote-debugging-port=9222 first, then set BROWSER_MODE=cdp.

If the CDP connection fails, the server now fails loudly (no silent fallback to a fresh browser) and tells you to start the debug-port browser first.

In CDP (or persistent-profile) mode the agent can also take over tabs you already have open: browser_list_pages shows them under "external tabs", and browser_adopt_page(ext_index) pulls one into the task (snapshot/click/read all work from then on). Adoption always requires confirmed=true — it hands the agent full read/write access to that page, including its login state.

Configuration (environment variables)

Every option can be overridden via BROWSER_-prefixed env vars:

评测/全功能使用：两个高危能力默认关闭（安全优先）。需要 browser_evaluate 时设 BROWSER_ALLOW_JS_EXECUTION=true，需要 browser_network_body 时设 BROWSER_ALLOW_NETWORK_BODY=true——冷启动跑基准/评测不开它们，对应子任务会被拒（这是设计，不是故障）。

Tools

34 tools: browser_navigate, browser_navigate_back, browser_snapshot, browser_click, browser_type, browser_hover, browser_press_key, browser_select_option, browser_upload_file (HITL-confirmed), browser_drag, browser_dialog_respond (dialogs are parked for agent/user decision; accept requires confirmed=true), browser_adopt_page (take over an already-open browser tab in cdp/persistent mode; HITL-confirmed), browser_read, browser_screenshot, browser_evaluate, browser_wait, browser_wait_stable, browser_wait_ms, browser_scroll, browser_scroll_to, browser_wait_navigation, browser_find (locate content by text → refs), browser_dismiss_popup, browser_list_pages, browser_switch_page, observability tools browser_console, browser_errors, browser_network, browser_perf, browser_network_body, plus 4 lifecycle tools: browser_tasks, browser_close_task, browser_list_sessions, browser_close_session.

Observability (debugging): console output, uncaught exceptions and request metadata are buffered per page from creation; browser_errors() returns a merged "JS exceptions + console.error + failed requests" view in one call. All three support a since cursor (omit = continue from last read, 0 = full) and limit paging.

Performance: browser_perf() returns FCP/LCP/CLS/INP, navigation timings and the 5 slowest resources. Response bodies can be fetched on demand with browser_network_body(seq) — off by default (BROWSER_ALLOW_NETWORK_BODY), every call gated by confirmed=true, hard char cap, and the body never enters the audit log.

Token cost: measured, not claimed

Same 10-step task, both servers at default config, metered at the JSON-RPC payload layer (cl100k tokens; harness + raw data in docs/bench/token-comparison.md, reproduce with bench/compare.py):

7.5x fewer tokens overall; 100x on repeated snapshots — the dominant cost in real agent loops (polling, multi-step forms, state confirmation).

Real-site benchmark (7 scenarios × 5-7 verifiable sub-tasks each: Baidu/Bing/DuckDuckGo search, Wikipedia reading, Hacker News, GitHub browsing — docs/bench/realworld.md): sub-task completion 37/42 vs 31/42, 18.2k vs 566.4k tokens (31x), wall-clock 107s vs 179s — one Wikipedia article snapshot alone costs pw-mcp ~252k tokens where nexus caps + diffs. Enterprise task suite (filter/sort, dashboard reading, KB answers, multi-step ordering, price comparison — docs/bench/enterprise-ops.md): 21/21 on all three servers (vs playwright-mcp and chrome-devtools-mcp); tokens 3.6k vs 5.1k vs 10.9k. At scale (**106 cases / 184 sub-tasks, three servers, see