JS Eyes

JS Eyes is a local-first browser capability and site-skill runtime for AI agents. A browser extension connects to a local JS Eyes server; CLI, MCP, and OpenClaw are peer host surfaces over the same protocol, policy engine, and Skill Runtime.

Treat {baseDir} as the root of this installed bundle. The optional OpenClaw adapter is {baseDir}/openclaw-plugin.

Use this Skill when

• The user wants to install or connect JS Eyes.
• An MCP or OpenClaw host cannot see JS Eyes tools.
• The browser extension remains disconnected.
• A local server, token, browser target, or security policy needs diagnosis.
• The user wants to discover, inspect, trust, enable, or call a JS Eyes Skill.
• An external V2 Skill needs to be linked without coupling it to OpenClaw.

Choose the host surface

Use the smallest surface that fits the request:

1. CLI — use js-eyes directly for server management, diagnostics, Skill lifecycle, and one-off Skill calls.
2. MCP — use @js-eyes/mcp-server for Codex, Claude, Cursor, VS Code, and other local MCP clients.
3. OpenClaw — load {baseDir}/openclaw-plugin only when OpenClaw-specific lifecycle and routing are required.

Do not install the OpenClaw adapter merely to use CLI or MCP.

Requirements

• Node.js 22 or newer.
• A supported Chrome, Edge, or Firefox extension.
• A local JS Eyes server, normally at http://localhost:18080.
• A shared server token unless anonymous compatibility mode was explicitly selected.

Keep the server bound to loopback unless the user has deliberately configured and secured remote access.

Standard standalone setup

Install the public CLI:

npm install -g js-eyes

Initialize a token, register the optional Native Messaging bridge, and start the server:

js-eyes server token init
js-eyes native-host install --browser all
js-eyes server start
js-eyes doctor

Install or load the browser extension, then use its popup to synchronize the server URL and token. If Native Messaging is unavailable, reveal the token only for the local user and paste it into the popup:

js-eyes server token show --reveal

Never place the token in documentation, logs, chat output, command arguments that will be shared, or a remote URL.

MCP setup

The MCP facade connects lazily to an existing JS Eyes server:

{
  "mcpServers": {
    "js-eyes": {
      "command": "npx",
      "args": ["-y", "@js-eyes/mcp-server"]
    }
  }
}

The default safe profile exposes browser status, tab, navigation, page-read, screenshot, and read-only Skill Runtime tools. Raw JavaScript, CSS injection, cookie access, file upload, and non-read Skill calls are not discoverable.

Use --tool-profile full only for a trusted MCP host when the requested task actually requires those capabilities. Server policy remains authoritative in both profiles.

Start diagnosis with:

1. browser_status
2. browser_list_clients
3. browser_list_tabs

When several extensions are connected, pass the exact clientId returned by browser_list_clients; do not select an ambiguous browser automatically.

OpenClaw setup

The ClawHub/bundle deployment uses the runtime packages and optional adapter shipped under {baseDir}:

1. Run npm install in {baseDir} with Node.js 22 or newer.
2. Resolve the OpenClaw config path in this order:
   • OPENCLAW_CONFIG_PATH
   • OPENCLAW_STATE_DIR/openclaw.json
   • OPENCLAW_HOME/.openclaw/openclaw.json
   • ~/.openclaw/openclaw.json
3. Merge the adapter path and enablement into the existing config.
4. Add js-eyes to tools.alsoAllow or the equivalent tool allowlist.
5. Restart or refresh OpenClaw.

Use this config shape without removing unrelated entries:

{
  "tools": {
    "alsoAllow": ["js-eyes"]
  },
  "plugins": {
    "load": {
      "paths": ["/absolute/path/to/js-eyes/openclaw-plugin"]
    },
    "entries": {
      "js-eyes": {
        "enabled": true,
        "config": {
          "serverHost": "localhost",
          "serverPort": 18080,
          "autoStartServer": true
        }
      }
    }
  }
}

JS Eyes registers one OpenClaw tool named js-eyes. It routes path-style actions such as:

• browser/get-tabs
• browser/list-clients
• browser/open-url
• browser/get-html
• skills/discover
• skills/plan-install
• skills/reload
• security/reload
• skill/<skillId>/<action>

Do not look for or invoke the removed pre-2.8 multi-tool family.

Verify OpenClaw mode in this order:

openclaw plugins inspect js-eyes
openclaw js-eyes status

Then call the js-eyes tool with action: browser/get-tabs and args: {}.

Safe defaults

The recommended host posture is:

• security.allowAnonymous=false
• security.allowRemoteBind=false
• security.allowRawEval=false
• security.enforcement=soft or strict
• a narrow security.egressAllowlist
• token authentication enabled

Ordinary status, tab, navigation, page-read, screenshot, and first-class browser operations should not require arbitrary JavaScript. If a specific legacy or site Skill requires raw execution, explain the risk and enable security.allowRawEval only with explicit user intent. Restart the server after changing that setting.

Review the effective posture with:

js-eyes security show
js-eyes doctor
js-eyes audit tail --lines 100

Policy or egress failures are decisions, not connectivity failures. Preserve the returned error code and pendingId, let the user review the request, and use the normal js-eyes egress or consent commands before retrying.

Skill Runtime V2

JS Eyes discovers a V2 Skill from static package.json and skill.manifest.json metadata without executing its entry module. The manifest declares compatibility, tools, schemas, risks, and capabilities.

Common lifecycle commands are:

js-eyes skills list
js-eyes skills inspect <id>
js-eyes skills enable <id>
js-eyes skills disable <id>
js-eyes skills link /absolute/path/to/skill
js-eyes skills unlink /absolute/path/to/skill
js-eyes skills trust <id>
js-eyes skills revoke <id>
js-eyes skills reload

Call the host-neutral runtime directly:

js-eyes skill call <id> <tool> --args '{"key":"value"}' --json

MCP exposes the same runtime through skill_list, skill_describe, and skill_call. OpenClaw routes the same tools through skill/<skillId>/<action>.

For external Skills under prompt or strict policy:

1. Link the directory.
2. Inspect its source, manifest, dependencies, declared permissions, and execution mode.
3. Trust it only after review.
4. Re-review it whenever its path, source digest, dependencies, manifest, capabilities, or execution mode changes.

Worker mode is a crash and reachability boundary, not an operating-system sandbox. Direct filesystem, process, and network access cannot be completely contained by a JavaScript Worker.

Native Messaging

Native Messaging lets an allowlisted extension read the local server URL and token without manual copy and paste:

js-eyes native-host install --browser all
js-eyes native-host status --browser all

Restart the browser or reload the extension after registration. If the popup reports token-missing, initialize the token before synchronizing. If the host is unavailable, use manual token entry instead.

This mechanism protects against ordinary external web pages and unlisted extensions. A compromised local device or malicious local process is outside its threat model.

Troubleshooting

Server unavailable

Run:

js-eyes status
js-eyes doctor

Start the server if necessary and confirm that the configured port is not already owned by an unrelated process.

Extension unavailable

Confirm:

• the correct extension is enabled;
• the server URL uses the configured local host and port;
• the extension and server share the same token;
• the browser was restarted after Native Messaging installation;
• at least one extension client appears in status output.

Authentication failed

Rotate or reinitialize the token only when the user understands that all connected clients must resynchronize:

js-eyes server token rotate

Never reveal the replacement token in shared output.

Browser target required

List clients and use the exact target ID. Do not guess when multiple browsers or profiles are connected.

Skill not found or disabled

Check js-eyes skills list, source directories, enablement, compatibility, and trust state. Use inspect before trust; do not bypass strict policy by silently switching it to legacy.

Skill input rejected

Read the schema returned by skill_describe or skills inspect. Correct the arguments instead of modifying the Skill to skip validation.

Raw execution refused

RAW_EVAL_DISABLED is the safe default. Prefer a first-class operation or read-only Skill tool. Enable raw execution only when the task requires it and the user accepts the additional authority.

Agent operating rules

• Diagnose before changing configuration.
• Repair existing configuration by merging; never overwrite unrelated hosts, plugins, tools, or security settings.
• Prefer CLI or MCP unless the task specifically requires OpenClaw.
• Prefer read-only Skill tools and the MCP safe profile.
• Ask for explicit intent before enabling a full profile, raw execution, cookie access, file upload, destructive tools, or administrative tools.
• Report the active host surface, connected browser target, and any policy decision when handing the setup back to the user.