Chrome Faithful Faithful control of your real, logged-in Chrome profiles. An MCP server + MV3 Chrome extension + authenticated localhost bridge that lets AI agents drive the Chrome that already holds your logins, extensions, and history. No copied profiles, no debug profile, no --remote-debugging-port, no Edge, no global mouse/keyboard automation. English · 简体中文 --- ## Why this exists Browser-control tools optimize for different jobs: | Approach | What you get | What you lose | |---|---|---| | Chrome DevTools MCP (Google) | Excellent DevTools, performance, and CDP workflows; Chrome 144+ can autoConnect to a running local browser with user approval | Chrome must already be running, and when several profiles are active Chrome chooses the default profile rather than accepting an exact profile name | | Playwright / Puppeteer MCP servers | Deterministic, isolated browsers that are ideal for CI and repeatable tests | Existing logins, extensions, history, and two-factor sessions are not present unless separately provisioned | | Extension-based MCPs (BrowserMCP, real-browser-mcp) | Control of an existing logged-in browser | A strong fit for live sessions; multi-profile setups may require separate server instances and ports, and normally expect Chrome to be running | Chrome Faithful focuses on exact-profile, multi-profile control with a fail-closed local bridge: - Exact multi-profile routing. Every profile registers with its exact profileName; duplicate registrations are rejected, so concurrent agents cannot interleave inside one profile. - Launches a closed profile. If the target profile — or all of Chrome — is closed, it starts the exact profile with ordinary Chrome and waits for the exact extension registration before reporting success. No --user-data-dir hacks. - Security depth. The bridge binds only 127.0.0.1 and requires a generated 256-bit secret. Bootstrap uses one-use tokens; sessions use scoped grants. Configuration is closed-schema and must live outside the source tree. Installers are transactional with SHA-256-verified, DPAPI-encrypted backups (Windows). - File upload the honest way. Files are injected as page File/DataTransfer objects — not DOM.setFileInputFiles, not an OS file chooser. - Media export without leaking URLs. chrome_page_asset streams page-exposed media using the tab's user agent, referer, and matching profile cookies; signed URLs, cookies, and headers never appear in MCP arguments or results. - Durable virtual-list capture. Scroll capture with asset parity, fail-closed manifests, exclusive cross-process locks, and resume that rewinds the tab through serialized wheel events — built for infinite-scroll feeds. - Works minimized. Locator waits/actions and screenshots use CDP focus emulation, so virtualized controls keep rendering even when the Chrome window is minimized or obscured. - Raw CDP when you need it, with an explicit trust boundary. chrome_cdp event reads redact Network headers, query strings, and post data; the bounded request/response projection actions reject sensitive selected fields. Its send action is deliberately unrestricted raw CDP and must be exposed only to a fully trusted MCP client: it can read authenticated page content, cookies, storage, tokens, URLs, and headers. - Structured verification handling. Multi-signal challenge detection that distinguishes resolved / pending-render / active challenge states, click-first solving for the common "click once and it passes" cases, and an honest handoff when a challenge needs a human — see Verification handling. - Codex-compatible JS API. src/agent-browser.mjs implements Codex's agent.browsers surface (tabs, locators, CUA, Playwright-style selectors, clipboard, dialogs, downloads) so JavaScript agents can use the same runtime. ## Architecture ┌─────────────┐ stdio ┌──────────────────────┐ ws://127.0.0.1 ┌─────────────────────────┐ │ MCP client │ ─────────► │ src/mcp-server.mjs │ ──────────────────► │ src/bridge-server.mjs │ │ (Claude, │ │ MCP tools (38) │ (Bearer secret) │ authenticated localhost │ │ Codex, …) │ └──────────────────────┘ │ multi-profile router │ └─────────────┘ └───────────┬─────────────┘ │ chrome.debugger ┌───────────────▼──────────────┐ │ MV3 extension in EACH exact │ │ profile (offscreen doc owns │ │ the WebSocket) │ └──────────────────────────────┘ - extension/ — MV3 extension loaded once per controllable profile. Uses chrome.debugger; an offscreen document owns the persistent WebSocket so MV3 service-worker suspension never drops the connection. - src/bridge-server.mjs — authenticated, localhost-only, multi-profile router with resilient failover. - src/chrome-profile-launcher.mjs — exact local Profile discovery and ordinary Chrome startup with bounded extension-registration confirmation. - src/mcp-server.mjs — the MCP tool surface (38 tools). - src/agent-browser.mjs — JavaScript agent.browsers compatibility adapter. - src/verification/ — challenge detection, hold state machine, handoff, overlay dismissal, humanized input, and the solve pipeline (checkbox / slider / click-first generic / capture-for-backend). - src/file-injection.mjs, src/page-asset.mjs, src/scroll-capture.mjs, src/scroll-asset-capture.mjs, src/network-request.mjs, src/network-response.mjs — the feature modules. - scripts/ — Windows installers, acceptance harnesses, live-test harness, and codex parity tooling. ## Safety model 1. A caller must select one exact metadata.profileName. 2. Duplicate live registrations for one profile name are rejected. 3. The bridge binds only 127.0.0.1 and requires a generated secret. 4. There is no fallback to a generic profile, port 9222, Edge, or UI automation. 5. If the target is disconnected, callers use chrome_profile_catalog / chrome_profile_start; a process start succeeds only after the exact extension profileName registers. 6. A live self-test must pass tabs and Runtime.evaluate before browser work. 7. Profile and tab failures are returned to the calling agent with no user-side console inspection required. chrome_cdp with action=send is outside the safe-projection boundary. It is equivalent to granting the MCP client DevTools access to the selected logged-in profile. Do not enable this server for untrusted clients or shared MCP hosts. See SECURITY.md for the full model and reporting policy. The extension's broad capabilities are intentional and visible: debugger provides DevTools-equivalent control; history, downloads, and clipboard permissions back their corresponding tools. Host access is limited to http://127.0.0.1/* for the local bridge. For deterministic, disposable CI browsers, use Playwright or Puppeteer instead. ## DSH first-class integration Chrome Faithful ships a first-party DeepSeek Harness bundle in packages/dsh-plugin-chrome-faithful/. It uses DSH's host-provided MCP client instead of duplicating the browser tools, so DSH gets the same exact-profile routing and security behavior as every other client. Supported baseline: @deepseek-ai/dsh 0.1.0-rc.6 and Node.js >=22.12.0. DSH remains an RC, so every newer RC requires a composition recheck. After the core and bundle packages are published, install into the intended profile: sh dsh plugin --profile web add @bpc-oss/dsh-plugin-chrome-faithful@0.4.0 The model sees stable names such as mcp__chrome_faithful__chrome_profiles. The bundle embeds no secret and passes AGENTOS_CHROME_CONFIG only when explicitly set. Initial configuration or resolution failures stop activation instead of leaving a silent zero-tool plugin. See the DSH bundle README for packaging, trust-boundary, and private-acceptance details. ### Local vision for text-only models chrome_visual_extract captures the requested exact-profile tab only when called, runs a local backend, and returns text JSON containing screenshot dimensions/SHA-256 plus OCR text, confidence, and normalized coordinates. It does not return or save the PNG. This makes the result useful to DSH models even though DSH 0.1.0-rc.6 drops MCP image content. The default backend is the shipped PP-OCRv5 mobile adapter. Chrome Faithful does not bundle or install Python, PaddleOCR, PaddlePaddle, OpenCV, NumPy, or model weights. Install those optional components yourself and configure both absolute local model directories so PaddleOCR cannot fall back to downloading weights: text CHROME_FAITHFUL_PYTHON=C:\Python311\python.exe CHROME_FAITHFUL_PPOCR_DET_MODEL_DIR=C:\Models\PP-OCRv5_mobile_det CHROME_FAITHFUL_PPOCR_REC_MODEL_DIR=C:\Models\PP-OCRv5_mobile_rec Live quality acceptance of the direct production MCP path detected 7/7 blocks, reached 99.43% raw and 100% non-whitespace character accuracy, and returned 0.9754 mean / 0.9379 minimum confidence with valid coordinates, reading order, and identical output across three 3.00--3.25 second calls on the accepted AMD Ryzen 9 9950X3D host. DSH model consumption of chrome_visual_extract output was not evaluated, and the optional VLM was not approved. See the acceptance report. CHROME_FAITHFUL_OCR_BACKEND may instead be a shell-free cli:["executable","arg"] specification or an exact http://127.0.0.1:<port>/... / http://[::1]:<port>/... endpoint. CHROME_FAITHFUL_VLM_BACKEND uses the same formats and is disabled by default; it can point to a user-operated SmolVLM2, Moondream, or compatible local adapter. Remote URLs, redirects, automatic downloads, and cloud fallback are rejected. Normalized OCR coordinates are hints for existing chrome_cua calls, not authorization to click. ## Quick start (Windows) Prerequisites: Node.js >= 22.12, Chrome, PowerShell (only the installers and the .cmd launcher are Windows-specific; the extension, bridge, and MCP server are platform-neutral). powershell npm ci --ignore-scripts 1. Load the extension in every Chrome profile you want agents to control: chrome://extensions → enable Developer mode → Load unpacked → select extension/. Note the 32-character extension ID and the loaded absolute path. 2. Create the bridge config outside the source tree, at %LOCALAPPDATA%\AgentOS\agentos-chrome-cdp\config.json, using config/local.example.json as the non-secret schema reference. The secret must be a generated 256-bit value, e.g.: powershell [Convert]::ToBase64String((1..32 | ForEach-Object { Get-Random -Max 256 })) The schema is closed: host (must be 127.0.0.1), port, secret, commandTimeoutMs, profileAliases — plus optional bridge/launcher overrides. The server refuses a config that lives inside the source tree. 3. Start the bridge: npm run bridge. 4. Register the MCP server with your client, pointing node at the absolute path of src/mcp-server.mjs (or use bin\invoke-chrome-cdp.cmd, which auto-starts the bridge). The in-repo .mcp.json uses paths relative to the repo root — that form works for Codex project configs; other clients generally want an absolute path. 5. Verify: call chrome_profiles, then chrome_selftest, then open a tab with chrome_tabs / navigate with chrome_tabs. For multi-profile client wiring, secret rotation, DPAPI-encrypted backups, and transactional rollback, the PowerShell installers automate it: powershell .\scripts\Install-AgentOsChromeExtension.ps1 -Target <absolute-loaded-extension-path> .\scripts\Install-AgentOsChromeCdp.ps1 -Clients @('CodeBuddy') -ExtensionId $ExtensionId -ExtensionPath $ExtensionPath -ChromeProfileDirectories $ProfileDirs -ChromeUserDataDir $ChromeUserData Run them without -Apply first — the default is a dry-run preview. ## MCP tools | Group | Tools | |---|---| | Profiles & sessions | chrome_profiles, chrome_profile_catalog, chrome_profile_start, chrome_selftest, chrome_session_v2 | | Tabs & navigation | chrome_tabs, chrome_session_v2 (finalize), chrome_page_event_v2 | | Interaction | chrome_playwright_v2, chrome_locator, chrome_cua, chrome_dom_cua_v2 | | Raw CDP & network | chrome_cdp, chrome_network_asset_v1 | | Capture & evidence | chrome_screenshot, chrome_visual_extract, chrome_cua_scroll_capture_v1/v2/v3, chrome_cua_scroll_capture_status_v1, chrome_cua_scroll_asset_capture_start/status/cancel_v2 | | Assets & content | chrome_page_asset, chrome_page_asset_v2, chrome_content_v2 (pdf/md/xlsx/csv/docx/pptx) | | Verification | chrome_verification_detect, chrome_verification_status, chrome_verification_resume, chrome_verification_solve, chrome_verification_solve_checkbox, chrome_verification_solve_slider, chrome_verification_capture, chrome_verification_dismiss_overlays | | Utilities | chrome_file_inject, chrome_history, chrome_clipboard | Notable behaviors: locator calls wait up to 30 s for visibility and are serialized per profile+tab; fill uses replacement semantics; chrome_locator accepts a zero-based index (-1 = last) for multi-match selectors; screenshots accept an optional document-coordinate clip and absolute savePath and still return the PNG. ## Verification handling Because Chrome Faithful drives your real profile, most bot checks never trigger. When a platform still presents a human-verification challenge, the verification module gives agents a structured loop instead of blind retries. Detection (chrome_verification_detect) classifies three real-world states: | State | What it means | Action | |---|---|---| | resolved | A token is already populated (e.g. invisible challenge completed) | Not a blocker — proceed | | active provider iframe (reCAPTCHA v2/v3, hCaptcha, Turnstile, GeeTest, vaptcha) | A visible challenge widget is present | Solve it | | pending-render | A widget container exists but its challenge iframe never rendered. Field data shows the usual root cause is a stale/expired session, not a network handshake stall | Refresh the session (sign out/in), trigger the page's real submit/verify button, reload-and-retry, or hand off | Static markers (the ubiquitous reCAPTCHA badge) are explicitly excluded, so a page that merely loads reCAPTCHA is never reported as a challenge. Solving (chrome_verification_solve) picks the strategy by type: 1. Checkbox / token wait — reCAPTCHA v2 / hCaptcha / Turnstile: click the visible challenge control (provider iframe center preferred) and poll the hidden response token until populated. For silent/interaction-only Turnstile with no visible checkbox, trigger the page's real submit/verify button with a JS click (btn.click() via page evaluation — locators time out when the button is off-screen or covered) so turnstile.execute() runs the actual challenge; the backend validates the real token, so never monkey-patch window.turnstile (fake/empty tokens are rejected, e.g. HTTP 422). 2. Humanized slider drag — GeeTest / slider: locate the handle, compute the target (track end or a backend gap offset), drag with a seeded bezier trajectory (monotonic x, jitter, ease-in-out delays), then verify acceptance. A gap behind the handle fails closed instead of dragging backwards. 3. Click-first generic — text-signal / unknown challenges: click the obvious "Verify you are human" / "验证" / "继续" button (or challenge checkbox) once, wait briefly for a token, and only then escalate. 4. Capture for backend — image-select / audio challenges: save the challenge image region and/or audio URL and submit them to an external OCR/ASR backend. Hold state machine (chrome_verification_status / chrome_verification_resume) — per-profile idle → challenge_detected → waiting_for_human → cleared with an auditable, bounded transition log. chrome_verification_solve clears the hold on success, hands off on failure, and rolls the hold back if the solver itself crashes. Humanized input — seeded bezier trajectories with jitter, monotonic-x slider drags, and ease-in-out timing (src/verification/input.mjs), deterministic and testable. Recognition backends are external and optional. Enable one via the AGENTOS_VERIFICATION_BACKEND environment variable, e.g. cli:python scripts/verification/captcha-backend-adapter.py (a reference JSON adapter for the Python faster-whisper / OCR / opencv stack; it prefers the Agent OS captcha connector when importable and falls back to standalone faster-whisper / ddddocr / tesseract / opencv otherwise) or an HTTP endpoint. Without a backend, detection, hold/resume, handoff, overlay dismissal, and humanized interaction all still work. Enabling a backend sends the configured process or endpoint a local capture path and/or a challenge audio URL plus the requested action; an HTTP endpoint may therefore transfer challenge data or credentials outside this project. Configure only an endpoint you trust and are authorized to use. Design: docs/superpowers/specs/2026-08-14-verification-handling-design.md Turnstile truth (field-verified). In a real submission workflow, a "widget rendered but challenge iframe never appears / token stays empty" state turned out to be a stale session, not an environmental dead-end: after signing out and back in on the same profile, the challenge rendered and completed normally. The working path is a real front-end click — trigger the page's actual submit/verify button (a JS btn.click() via page evaluation works even when the button is off-screen) so turnstile.execute() runs the real challenge, then let the challenge-complete callback submit the real token. In the field case the platform's submit endpoint validated that token strictly (fake/empty → HTTP 422), and monkey-patching window.turnstile (getResponse / render-with-immediate-callback / hidden-input injection) never produces one. Escalation order for pending-render: refresh session → trigger real submit → reload-and-retry → hand off. ## Live testing scripts/verification/live-tests/ contains reproducible harnesses that drive real Chrome profiles through the compliant bridge channel (task tabs only; they are closed after each run): - live-verification-test.mjs [url] [profileName] — generic detect → solve → re-detect loop against any URL. - live-cf-test.mjs [profileName] — Cloudflare Turnstile with the official test sitekeys (1x00000000000000000000AA always-pass, 3x00000000000000000000FF forced interactive) plus a click-to-pass simulation fixture. Serve fixtures with python -m http.server 18999 --directory scripts/verification/live-tests. - cf-diagnostic-probe.mjs [profileName] — dumps widget markup / iframe / window.turnstile state for the "widget rendered but challenge iframe missing" (pending-render) state. - final-regression.mjs [profileName] — badge-only pages must not be detected; click-to-pass must still solve. ## JavaScript integration js import { startBridge, createAgent } from "./src/index.mjs"; const bridge = await startBridge(); const agent = createAgent(bridge.router); const targets = await agent.browsers.list(); const browser = await agent.browsers.get(targets[0].id); const tab = await browser.tabs.new(); await tab.goto("https://example.com/"); ## Codex compatibility src/agent-browser.mjs implements the Codex agent.browsers surface. Parity is pinned mechanically: compat/ holds a repository-authored functional surface contract, its adapter-map SHA-256, and the adapter map. It does not redistribute bundled product documentation. npm run check:parity and test/codex-parity-contract.test.mjs fail if any contract member is missing, stubbed, or extra. See compat/README.md and docs/CODEX_PARITY.md. The internal identifiers agentos-chrome-cdp, AGENTOS_CHROME_CONFIG, and the existing AgentOS configuration path are retained for upgrade compatibility; the public display name is Chrome Faithful. ## Testing text npm run check # static gates (structure, JSON validity, generic boundary) npm test # mock/unit tests, incl. security contract tests npm run check:parity # Codex agent.browsers parity contract npm run build:extension Installer transaction tests (Windows): pwsh -NoProfile -File test/installer-transactions.test.ps1. Static checks and mock tests are necessary but not sufficient. Release acceptance additionally requires two concurrently connected real profiles, per-profile selftest, background tab navigation, locator click/fill, raw CDP, screenshot, history, clipboard round-trip with restoration, dry page-File injection, reconnect, and proof that only acceptance-owned tabs were closed — driven by scripts/live-acceptance.mjs, scripts/differential-acceptance.mjs, and the live-test harness above. ## Documentation - SECURITY.md — safety model and vulnerability reporting - CONTRIBUTING.md — development workflow - docs/CODEX_PARITY.md — Codex parity design - packages/dsh-plugin-chrome-faithful/ — first-party DSH bundle - docs/superpowers/specs/ — design specs (profile launch, resilient bridge ownership, verification handling) - skills/control-chrome-cdp/SKILL.md — agent-facing operating skill - THIRD_PARTY_NOTICES.md — bundled third-party code ## Status Experimental. Windows-first: the installers, DPAPI backups, and .cmd launcher are Windows-only; the extension, bridge, and MCP server are platform-neutral Node.js and should run anywhere Chrome does, but only Windows is exercised today. The bridge controls your real logged-in profiles — review the safety model, use exact profiles, and never paste your bridge secret. ## License MIT. Bundled runtime code and development-only verification tools retain their own MIT, Apache-2.0, or ISC terms — see THIRD_PARTY_NOTICES.md.