About Weavetab
Weavetab is a local automation engine for AI agents to control Chromium browsers and Electron apps through CDP. It represents pages as compact semantic maps with change diffs, and includes credential masking and local security policies.
Key features
- 01Control browsers and Electron through CDP
- 02Create semantic maps with change diffs
- 03Interact with and navigate pages
- 04Inject and mask credentials server-side
- 05Apply domain policies and role permissions
Use cases
- 01Automate web forms and navigation
- 02Run tasks in Electron applications
- 03Keep credentials out of agent context
- 04Inspect page changes and extract information
MCP support
The Weavetab MCP server exposes browser and Electron automation tools over stdio or a WebSocket bound to localhost. Clients can call tools for page inspection, interaction, navigation, screenshots, network access, and storage.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
Tools
42Saved tool list · updated 2w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| app | string | No | Electron app name configured in config.json under 'electron'. |
| full | boolean | No | Set to true to return a full markdown table (heavier context) instead of the lite list. |
| delta | boolean | No | Diff mode: return ONLY elements that changed since the last browser_map call (added, removed, or modified). Diffs across time — distinct from prune. Cuts prompt tokens by up to 97% on consecutive calls. |
| force | boolean | No | Force a full DOM map rebuild even if loop detection says nothing changed. Bypasses the map skip guard and resets loop strikes. |
| prune | boolean | No | Semantic DOM pruning: compress the current snapshot before returning. Drops hidden elements with no interactive descendants and not referenced by any toggle trigger (aria-controls etc.). Replaces SVG icons with [SVG Icon: label] tokens. Distinct from delta:true — delta diffs across time, prune compresses this snapshot once. Combine both for maximum token efficiency. |
| query | string | No | Filter returned elements by label or role substring. Useful on dense dashboards. |
| scope | enum | No | Restrict parsing to a semantic page area to save tokens. Useful on dense dashboards.navigation | main_content | sidebar | form | modal | utility |
| tab_id | string | No | Read a specific background tab instead of the active one. |
| visual | any | No | Set-of-Marks visual grounding. true = always inject numbered badges + return annotated screenshot. "auto" = only activate when DOM interactive count < 5 (canvas/WebGL/Figma pages). After receiving marks, use browser_click({ mark: N }) to interact with numbered elements. |
| ax_tree | boolean | No | Bypass the extension entirely and read raw Chrome Accessibility Tree. Use this if Shadow DOMs hide elements. |
| show_overlays | boolean | No | Default false. Debug only — draws colored element borders via extension (ugly). The blue edge glow is the standard visual indicator, shown automatically during reading states. Only enable this for debugging element boundaries. |
Details
From the project README
The open automation engine for AI agents.
Control live browsers and desktop apps locally with instant execution.
Tool Reference · Documentation · Architecture · Support · Security
Weavetab is the open automation engine for AI agents. It connects models directly to live Chromium browsers and Electron desktop applications via native Chrome DevTools Protocol (CDP) sockets — no WebDriver binaries, no cloud relay, zero telemetry.
Actions execute with sub-10ms latency. Web pages are translated into agent-readable semantic maps with turn-by-turn diffing that reduces token usage by up to 95%. Built-in security enforces local domain policies, role-based access control, and client-side credential injection.
Built and architected by fy2ne.
Why not Playwright or Puppeteer?
The standard approach — Playwright over MCP — works. But it carries fundamental architectural costs that compound at scale.
| Playwright / Puppeteer | Weavetab | |
|---|---|---|
| Protocol layer | WebDriver BiDi or CDP via browser binary | Raw CDP WebSocket, Layer 0 |
| Tool execution latency | 50–300 ms (binary + IPC overhead) | 3–8 ms |
| Desktop app control | Browser only | Chrome + any Electron app (Discord, VS Code, Figma…) |
| DOM representation | Full HTML dump to LLM | Semantic Map 3.0 with delta diffing |
| Token cost per map | ~2,000–8,000 tokens | ~80–400 tokens (95% reduction via delta) |
| CAPTCHA | API service required | Zero-API via CDP mouse physics |
| Credential safety | Secrets passed in plaintext | Blind Injection — never in LLM context |
| Extension access | Limited | Full Chrome Extension API surface |
Architecture
MCP Client (Claude / Cursor / Antigravity / Copilot / Windsurf)
|
| JSON-RPC 2.0 (stdio or WebSocket 127.0.0.1:3000)
v
+-------------------------------------------------------------+
| Weavetab MCP Server |
| |
| +-----------------+ +---------------------------------+ |
| | Security Layer |-->| Intelligence Engine | |
| | Rate · RBAC | | Map 3.0 · Delta · Memory | |
| | Policy · Audit | | Pruner · Framework Detector | |
| +-----------------+ +----------------+----------------+ |
| | |
| +----------v----------------+ |
| | CDP Bridge | |
| | Ghost Cursor · Typing | |
| | App Router · Loop Guard | |
| +----------+----------------+ |
+-------------------------------------------------------------+
| CDP WebSocket (127.0.0.1)
+---------------------------+-----------------------------+
v v v
Chrome / Edge VS Code / Discord Your Electron App
(any tab) (Electron apps) (local dev)
No cloud. No external process. Both the CDP WebSocket connection and the optional MCP WebSocket daemon (--ws) are strictly bound to 127.0.0.1 with shared-secret authentication (~/.weavetab-system/ws-secret) and Origin/Host header verification.
Quick Start
npx -y @weavetab/mcp
Add to your MCP client config:
{
"mcpServers": {
"weavetab": {
"command": "npx",
"args": ["-y", "@weavetab/mcp"]
}
}
}
| Client | Config path |
|---|---|
| Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json |
| Claude Code | claude mcp add weavetab -- npx -y @weavetab/mcp |
| Cursor | .cursor/mcp.json |
| Antigravity | ~/.gemini/antigravity/mcp/weavetab/ |
| VS Code Copilot | .vscode/mcp.json |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
Core Technology
Semantic Map 3.0
The primary bottleneck in browser-based agents is DOM representation. Full HTML dumps are noisy, large, and structurally meaningless to language models. Weavetab's browser_map runs an in-browser CDP walker that produces a compact semantic map:
╔═ ACTIVE DIALOG: "Confirm Payment" ═╗
[w:3] button "Confirm" [w:4] button "Cancel"
┌ "Pro Plan — $49/mo"
[w:12] button "Upgrade" [w:13] button "View Details"
[w:14] input val:"[email protected]" ph:"Email address" FOCUSED
[w:20...w:38] 18 nav links: "Dashboard" ... "Sign out"
Key properties:
- Ref IDs (
w:N) are stable within a session and used by all interaction tools - Container binding — buttons and inputs are grouped under their parent card, row, or section header
- Modal hoisting — open dialogs are isolated at the top so agents cannot miss them
- Link collapsing — repeated nav/footer runs of 4+ links compress into a single token line
- Semantic delta — on re-reads, only changed nodes are transmitted (
~modified,+added,−removed), cutting re-read cost by up to 95%
Ghost Cursor
Clicks are not dispatched as synthetic DOM events. They are delivered as a physical sequence of Input.dispatchMouseEvent CDP calls along a Bezier curve trajectory with randomized acceleration, natural jitter, and hold durations. Anti-bot fingerprinting that flags event-based automation does not trigger.
browser_pointer supports hover_chain for multi-level menu traversal in a single tool call:
{
"hover_chain": ["File", "Export As", "PNG"],
"click_final": true
}
Blind Credential Injection
browser_type_secret resolves secrets server-side from ~/.weavetab/secrets.json. The resolved value is never included in the MCP response, never appears in DOM snapshots, and is permanently masked as [REDACTED] in all audit logs. The LLM context never sees the credential.
weavetab secrets set GITHUB_TOKEN
# Stored. Never echoed, never logged, never surfaced to the agent.
Electron & Desktop App Automation
Any tool accepts an optional app parameter that routes the command to a registered Electron process instead of a browser tab:
{ "app": "discord", "ref": "w:24", "text": "Hey, deploying now" }
Register apps in ~/.weavetab/config.json:
{
"electron": {
"discord": "C:\\Users\\<user>\\AppData\\Local\\Discord\\Discord.exe",
"vscode": "C:\\Users\\<user>\\AppData\\Local\\Programs\\Microsoft VS Code\\Code.exe",
"figma": "C:\\Users\\<user>\\AppData\\Local\\Figma\\Figma.exe"
}
}
Weavetab auto-attaches via CDP to the registered process. No code changes required in the target application.
Zero-API CAPTCHA Solving
browser_captcha solves Cloudflare Turnstile, reCAPTCHA v2/v3, GeeTest, and hCaptcha by orchestrating CDP mouse events that replicate human behavioral signatures — velocity curves, micro-pauses, scroll inertia — with no third-party API, no external service, and no API key required.
Tool Suite
41 tools across 9 categories. Full parameter reference: docs/tools/tools.md
| Category | Tools |
|---|---|
| Vision & DOM | browser_map browser_find browser_scrape browser_inspect browser_snapshot browser_screenshot browser_console browser_extract_design |
| Input & Interaction | browser_click browser_type browser_type_secret browser_key browser_pointer browser_select browser_fill browser_upload browser_scroll |
| Navigation & Environment | browser_navigate browser_tabs browser_detect browser_viewport browser_emulate |
| Mission & Automation | browser_burst browser_automation browser_recording browser_pattern_learn |
| Safety & Human Oversight | browser_dialog browser_thoughts browser_clipboard browser_wait |
| Autonomous CAPTCHA | browser_captcha |
| Network & Storage | browser_network_intercept browser_storage browser_cookies browser_checkpoint browser_performance |
| Office Documents | browser_office |
| Canvas & Eval | browser_eval browser_pdf browser_canvas browser_highlight |
Security
- Transport: Both CDP WebSocket and MCP WebSocket daemon (
--ws) are strictly bound to127.0.0.1with 256-bit shared-secret Bearer/handshake authentication andOrigin/Hostvalidation. No listener on external interfaces. - Policy Engine: Per-domain
allow/read-only/deny/auditrules under the top-levelpoliciesarray in~/.weavetab/config.json(blocks*.bank.*andpaypal.comout of the box) - RBAC: Five roles —
operator(default),viewer(least-privilege read-only),automation,system,admin— enforce tool-level permission gates - Plugin Trust & Manifests: 3-tier verification (
@weavetab/*official, verified registry allowlist, and mandatory interactive confirmation +weavetab.jsonpermission manifest inspection for unverified packages) - Blind Injection: Credentials resolved server-side, masked in all outputs and audit logs
- Loop Breaker: Strike-based stuck-loop detector halts runaway autonomous sequences
- Audit Log: All tool invocations, policy decisions, and errors written to
~/.weavetab-system/logs/mission_<timestamp>.log
Security vulnerabilities: [email protected] (do not use public issues).
CLI Reference
weavetab # Start in stdio mode (default)
weavetab --ws --port 3000 # WebSocket daemon mode (bound to 127.0.0.1)
weavetab config # Interactive configuration wizard
weavetab secrets set KEY_NAME # Store a secret (never echoed)
weavetab secrets list # List stored secret keys (not values)
weavetab plugin add <name> # Install a verified/confirmed plugin
weavetab plugin list # List installed plugins
Configuration file: ~/.weavetab/config.json
{
"browserType": "chromium",
"headless": false,
"extension": true,
"maxActionsPerMinute": 60,
"role": "operator",
"electron": {
"discord": "...",
"vscode": "...",
"antigravity": "..."
}
}
Extensibility
Weavetab exposes a typed Plugin SDK for authoring custom MCP tools, HUD overlays, and tool lifecycle hooks:
npx @weavetab/sdk init my-plugin
See docs/PLUGINS.md for architecture, manifest spec, and SDK API reference.
License
AGPL-3.0-only. Client connections over stdio or WebSocket do not extend copyleft obligations to your client application or models.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.