Unbrowse (OpenClaw Plugin & Skill Ecosystem)

Open source API reverse engineering for OpenClaw.

Unbrowse captures real browser network traffic and turns it into reusable agent skills.

Security note: core capture/replay is local by default. Data is kept on your machine unless you explicitly publish. See SECURITY.md.

Why this project exists

Agents that drive websites through browser automation are slow and brittle:

• full browser startup
• DOM waits and selector drift
• repeated render/parse loops for every action

Unbrowse short-circuits that path:

• first run: observe with browser to capture real traffic
• later runs: call the same behavior directly through inferred endpoint contracts

Result: faster execution and fewer UI-shape failures in action-heavy workflows.

This is the practical “agentic web” move: agents should execute capability contracts, not scrape screen pixels.

Quick Start (works without config)

1) Install

openclaw plugins install @getfoundry/unbrowse-openclaw
openclaw gateway restart

2) Capture a local skill

unbrowse_capture { "urls": ["https://example.com"] }

3) Replay immediately

unbrowse_replay { "service": "example" }

What’s needed for each feature

Feature	What you need
Capture APIs	nothing extra
Generate artifacts	nothing extra
Replay local	nothing extra
Browse/login capture	browser session
Search shared skills	nothing extra
Publish/install sharing	marketplace route (optional)
Wallet payout flows	inactive in this repo

Payments are not enabled. Wallet tooling and config fields may exist, but settlement remains inactive.

System layout

• packages/plugin: local capture, inference, local replay, local artifact writes.
• server + packages/web: optional publish/search/install contracts, shared execution routes, validation merge surface.

The same command line can stay local-only indefinitely; marketplace participation is opt-in.

The 100x story (practical framing)

Browser-automation path

• launch browser
• render page
• wait for JS hydration
• inspect DOM / click and type
• scrape rendered output

Unbrowse path after learn

• execute the captured endpoint contract (GET, POST, etc.)
• reuse inferred auth/session context
• run deterministic request/response flows

The gain is not a “small optimization.” It is often the difference between workflows that stall and ones that feel immediate.

How it works

1. Capture
   unbrowse_capture records network traffic from your session.
2. Learn / infer
   Endpoints, methods, auth styles, and request shapes are inferred.
3. Generate
   Skill artifacts are written locally:
   • ~/.openclaw/skills/<service>/SKILL.md
   • scripts/
   • optional references/
   • optional auth.json
4. Replay
   unbrowse_replay executes locally first via browser/node mode.
5. Publish (optional)
   unbrowse_publish proposes shareable artifacts; server validates + merges and exposes searchable IDs.
6. Install / execute (optional)
   unbrowse_search can install discovered IDs into your local flow.

Core data and runtime boundary

Area	Local runtime	Remote/runtime contracts
Capture	✅ local sessions	❌ not required
Replay	✅ local artifacts	⏯️ optional backend path
Auth/session storage	✅ local-first	✅ explicit placeholder boundaries
Search/discover	❌	✅
Execution contracts	local fallback exists	optional through backend endpoint IDs

payments are intentionally not active in this repository.

Why local-first + publish-at-will

• private work should not depend on the index
• private credentials stay local
• shareable behavior is still possible after publish
• reusable ecosystem growth without forcing central lock-in

Plugin configuration

Set your plugin block in OpenClaw as plugins.entries.unbrowse-openclaw.config.

{
  "plugins": {
    "entries": {
      "unbrowse-openclaw": {
        "config": {
          "browserPort": 8891,
          "browserProfile": "",
          "allowLegacyPlaywrightFallback": false,
          "skillsOutputDir": "~/.openclaw/skills",
          "autoDiscover": true,
          "autoContribute": true,
          "enableAgentContextHints": false,
          "publishValidationWithAuth": false,
          "skillIndexUrl": "https://index.unbrowse.ai",
          "creatorWallet": "",
          "skillIndexSolanaPrivateKey": "",
          "credentialSource": "none",
          "enableChromeCookies": false,
          "enableDesktopAutomation": false,
          "telemetryEnabled": true,
          "telemetryLevel": "standard"
        }
      }
    }
  }
}

Option	Default	What it does
browserPort	OPENCLAW_GATEWAY_PORT / CLAWDBOT_GATEWAY_PORT / config.gateway.port / 18789, then +2	Browser bridge port for session tools
browserProfile	open-claw root default profile	Browser context profile selector
allowLegacyPlaywrightFallback	false	Enable Playwright fallback on direct browser flow
skillsOutputDir	~/.openclaw/skills	Skill write location
autoDiscover	true	Reuse previously captured skill hints automatically
autoContribute	true	Auto-contribute publish flow when enabled
enableAgentContextHints	false	Add captured-session context hints
publishValidationWithAuth	false	Validate publish calls with auth context
skillIndexUrl	UNBROWSE_INDEX_URL / https://index.unbrowse.ai	Marketplace/search base URL
creatorWallet	saved wallet or UNBROWSE_CREATOR_WALLET	Creator payout wallet
skillIndexSolanaPrivateKey	UNBROWSE_SOLANA_PRIVATE_KEY	Payer private key for marketplace operations
credentialSource	UNBROWSE_CREDENTIAL_SOURCE or none	Credential lookup: none, vault, keychain, 1password
enableChromeCookies	false	Allow Chrome cookie read path
enableDesktopAutomation	false	Enable unbrowse_desktop
telemetryEnabled	true	Send usage telemetry to index
telemetryLevel	standard	Telemetry payload detail: minimal, standard, debug

Plugin env helpers:

• OPENCLAW_GATEWAY_PORT / CLAWDBOT_GATEWAY_PORT
• UNBROWSE_INDEX_URL
• UNBROWSE_CREATOR_WALLET
• UNBROWSE_SOLANA_PRIVATE_KEY
• UNBROWSE_CREDENTIAL_SOURCE
• OPENROUTER_API_KEY
• OPENAI_API_KEY

Local override:

• ~/.openclaw/unbrowse/telemetry.json with { "enabled": boolean, "level": "minimal" | "standard" | "debug" } takes precedence over telemetryEnabled / telemetryLevel.

What isn’t documented (and why)

We intentionally treat these as black-box/internal:

• ranking internals
• settlement routing internals
• partner execution topology

The docs focus on:

• tool contracts
• observable inputs/outputs
• merge/conflict behavior
• security boundaries

Local and shared command map

• Capture / learn
  • unbrowse_browse
  • unbrowse_capture
  • unbrowse_learn
  • unbrowse_login
  • unbrowse_auth
• Skill/runtime execution
  • unbrowse_replay
  • unbrowse_skills
• Marketplace
  • unbrowse_search
  • unbrowse_publish
• Automation helpers
  • unbrowse_do
  • unbrowse_desktop
• Workflows
  • unbrowse_workflow_record
  • unbrowse_workflow_learn
  • unbrowse_workflow_execute
  • unbrowse_workflow_stats
• Wallet/config
  • unbrowse_wallet (currently inactive for settlement path in this repo)

Installation options

Recommended

openclaw plugins install @getfoundry/unbrowse-openclaw

Manual JSON install

Add to your OpenClaw config (or equivalent runtime config file):

{
  "plugins": {
    "entries": {
      "unbrowse-openclaw": {
        "enabled": true
      }
    }
  }
}

Then:

openclaw gateway restart

Run from source

git clone https://github.com/lekt9/unbrowse-openclaw ~/.openclaw/extensions/unbrowse-openclaw
cd ~/.openclaw/extensions/unbrowse-openclaw
npm install
openclaw gateway restart

Next reads

1. docs/AGENTIC_WEB.md
2. docs/ARCHITECTURE.md
3. docs/INTEGRATION_BOUNDARIES.md
4. docs/CONTRIBUTOR_PLAYBOOK.md
5. docs/PURPOSE.md