User repository on GitHub
BrowserClaw - Control your everyday Chrome browser from AI agents, without losing logins or focus
- Forks
- 3
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
BrowserClaw
Chrome MV3 extension plus a local MCP server exposing 50 browser tools; its chrome_act_toward_goal micro-loop uses TypeSafe Jev System One or a local decider to pick the next action on an 85%+ pruned, Shadow-DOM-pierced DOM tree, dispatching trusted CDP events in logged-in Chrome without stealing focus.
- Stars
- 32
- Language
- TypeScript
- License
- AGPL-3.0
- Published by
- Community
- Last verified
- Oct 2026
About this project
BrowserPaw Control your everyday Chrome browser from AI agents, without losing logins or focus. Project Map · Tool Reference (50) · Client Config · Chinese (zh-CN) · AGPL-3.0 license · Releases. The Backstory: Why I Built BrowserPaw. Traditional browser automation frameworks and LLM browser-use agents were designed for isolated sandbox testing, not for real daily desktop workflows. Every time I tried to let an AI agent assist me in my real browser, I kept hitting the same six frustrating roadblocks:
Quoted from the project's README. Read the full README.
- Languages
- TypeScript
- Platforms
- MCPChrome extensionBrowserAPI
- Capabilities
- Open sourceRuns locallyRequires API key
Highlights
- Ships 50 MCP tools across 8 categories, defaulting to a 14-tool core profile to limit token use.
- chrome_read_dom prunes non-interactive and occluded nodes; the project reports over 85% node reduction on 1,000+ node pages.
- The chrome_act_toward_goal loop is bounded: at most 60 steps in Jev mode (10 by default), intercepting 14 destructive action keywords.
- Clicks and keystrokes are dispatched as trusted CDP events (isTrusted: true), so React/Vue/Angular and Shadow DOM handlers fire.
- Jev mode runs against TypeSafe cloud inference (TYPESAFE_API_KEY) or a local decider-2b model on port 8009 with CUDA or CPU fallback.
Watch out
AGPL-3.0 license; the Jev cloud mode needs a TypeSafe API key, and the README still brands the project BrowserPaw in several places and links to a browserpaw repository path.
Questions about this entry
Short answers drawn from the linked project or page.
No. It runs as a Chrome MV3 extension plus a local native messaging bridge inside the Chrome you already use, so Google, GitHub, and SSO sessions remain. It dispatches trusted CDP events in background tabs without taking focus or requiring manual approval popups.
The native server perceives the pruned page, asks Jev (or a local decider model) for the next action toward the natural-language goal, acts, and verifies, without an MCP round-trip per step. Ambiguous or low-confidence moves escalate back to the calling agent.
Three routes. The recommended one asks an agent to read INSTALL.md in the repository and follow it, which compiles the project and registers the OS native host, then you load app/chrome-extension/.output/chrome-mv3 unpacked. A prebuilt extension ZIP plus skill ZIP are on Releases, but the extension archive does not include the native messaging server, so the bridge still needs a clone and a build. From source it is `pnpm install`, `pnpm build`, then `node app/native-server/dist/scripts/register-dev.js`.
Off disables the micro-loop in the extension popup and hides chrome_act_toward_goal from the tool list, rejecting direct calls until you authorize it. Local runs an offline decider-2b model on port 8009 with NVIDIA CUDA acceleration or CPU fallback, hot-loading once weights are downloaded. Remote uses TypeSafe Jev System One cloud inference with TYPESAFE_API_KEY, an optional custom base URL and a custom fine-tuned modelId, and a broken configuration returns diagnostic logs instead of quietly falling back to heuristics.
It is bounded: at most 60 steps in Jev mode with 10 as the default, truncated to 5 in heuristic fallback. It intercepts 14 destructive action keywords such as pay, delete and submit, and escalates ambiguous or low-confidence decisions back to the calling agent with the candidate elements attached.
chrome_read_dom strips non-interactive and occluded nodes and indexes what remains, which the project reports cuts node count by over 85% on pages of 1,000 or more nodes, with a fast snapshot mode targeting 10-30 ms and 15 KB or less. Card flattening collapses repetitive feed cards into one-line summaries and viewport virtualization folds off-screen list items into count placeholders.
50 tools across 8 categories, defaulting to a 14-tool core profile for everyday sessions to avoid token bloat and decision paralysis. Full exposure is enabled with CHROME_MCP_TOOL_PROFILE=full or by activating tools dynamically through chrome_tool_docs.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Show HN: Agent Chaperone screens tool calls and results with Jev before the agent sees them.
sepehrsafari5 pointsShow HN: Agent Chaperone – Screen AI agent tool calls and results with JevRead the thread on Hacker NewsX: An X timeline labeler that flags engagement bait
X: Unclutter: an ad and slop blocker that runs on Jev
X: dmx.to reshapes the timeline in plain language
More Jev resources like this
18GitHub stars
MCP server that lets Claude Code, Cursor and other clients drive Chrome over the DevTools Protocol, exposing an accessibility-tree menu of stable refs. Its example Jev loop has Jev pick browser actions from that menu, and the README reports about a third fewer tokens and a quarter fewer tool calls in a blind benchmark.
Jev Agents & MCP#community#mcp#browser-automation
Community
Silbercue/public-browserTypeScript
Silbercue/public-browserTypeScript45GitHub stars
Five typed tools — classify, check, score, rank, and ask — plus a skill, wired into Claude Code, Codex, Pi, and OpenCode in one setup command.
Jev Agents & MCP#community#typescript#mcp
Community
FrancoisChastel/jev-codeTypeScript
FrancoisChastel/jev-codeTypeScript22GitHub stars
An MCP proxy and hooks adapter that screens an agent's tool calls before they run and the tool results before the agent reads them, with thresholds in a policy file.
Jev Agents & MCP#community#mcp#security
0GitHub stars
MCP server and Claude Code plugin that files each captured note, link or meeting line with one Jev request: kind, project, action and duplicate questions answered in parallel. Routes to a project only at 0.45 or above, marks a duplicate only at 0.70 with a matching item, and stores text verbatim in append-only JSONL.
Jev Agents & MCP#community#mcp#note-taking
Community
dharun-cohere/jev-sortwellTypeScript
dharun-cohere/jev-sortwellTypeScript9GitHub stars
Local code search for coding agents: ripgrep and BM25 shortlist function-level chunks, Jev Noul relevance questions run across parallel requests, and accepted excerpts return to Codex, Claude Code, OpenCode or Pi through MCP or the `oko ask` command.
Jev Agents & MCP#community#mcp#code-search
Community
bartlomein/okoRust
bartlomein/okoRust32.5kGitHub stars
A Claude Code configuration CLI shipping Jev modules for model routing, skill suggestion, tool guardrails, auto-mode gates, and sandbox hooks.
Jev Agents & MCP#community#typescript#coding-agent
Community
davila7/claude-code-templatesPython
davila7/claude-code-templatesPythonFrom the Jev community
Posts from builders shipping with Jev right now.