Cgw

Русская версия

Chrome extension + MCP server for connecting AI agents to corporate systems via browser session. Capture endpoints from your real work, curate them into skills, and let the agent accumulate knowledge across sessions.

Installation Guide | Creating Custom Skills | Agent Knowledge Layer | Security

How It Works

┌──────────────────────────────────────────────────────────────┐
│                                                              │
│  AI Agent (OpenCode, Cursor, Claude Code)                    │
│       │                                                      │
│       │ POST /mcp (Bearer token)                             │
│       ▼                                                      │
│  ┌──────────┐    WebSocket     ┌───────────────────────┐     │
│  │ cgw_mcp  │◄────────────────►│  Chrome Extension     │     │
│  │ :9877    │                  │                       │     │
│  │          │                  │  • chrome.cookies     │     │
│  │ Tokens:  │                  │  • chrome.webRequest  │     │
│  │ • agent  │                  │  • fetch() w/ session │     │
│  │ • ext    │                  │  • Skill management   │     │
│  └──────────┘                  └───────────┬───────────┘     │
│                                            │                 │
│                                            │ fetch() + cookies/auth
│                                            ▼                 │
│                                   Corporate APIs             │
│                                   (Jira, Mattermost, ...)    │
└──────────────────────────────────────────────────────────────┘

Key advantage: the extension reuses your existing browser session for authorization. No need to store passwords, API keys, or configure OAuth — if you're logged into a corporate system in Chrome, the extension automatically uses that session.

Components

Component	Description	Location
Chrome Extension	Skill management, request execution via browser session	extension/
cgw_mcp	MCP server (HTTP + WebSocket daemon), bridge between agent and extension	cgw_mcp/
Presets	Ready-made skill sets for popular systems	presets/

Quick Start

1. Install the extension

1. Open chrome://extensions
2. Enable Developer mode
3. Click Load unpacked → select the extension/ folder

2. Install and run cgw_mcp

cd cgw_mcp
npm install

As a daemon (autostart):

# Linux / macOS
./install.sh

# Windows (PowerShell)
.\install.ps1

Manually:

node index.js

On first run, a config file is created at ~/.corpgateway/cgw_mcp.json:

{
  "port": 9877,
  "token": "abc123...",
  "extensionToken": "def456...",
  "mcpInstructions": "..."
}

3. Connect the extension to cgw_mcp

1. Extension icon → ⚙ Settings
2. In the MCP Connection section:
   • Instance name: anything (e.g. "Chrome Work")
   • MCP server URL: http://localhost:9877
   • Extension token: copy extensionToken from ~/.corpgateway/cgw_mcp.json
3. Click Save
4. In the extension popup, click ⚡ Connect
5. The extension icon turns colored — connection established

4. Import skills

1. Extension settings → Import
2. Select a file from presets/ (e.g. jira.json)
3. Replace URL placeholders with your actual system addresses

5. Connect the AI agent

Add to your agent config (opencode.json, .cursor/mcp.json, etc.):

{
  "mcp": {
    "corp": {
      "type": "remote",
      "url": "http://localhost:9877/mcp",
      "headers": {
        "Authorization": "Bearer <token from cgw_mcp.json>"
      }
    }
  }
}

MCP Tools

The agent receives the following meta-tools:

Tool	Description	Parameters
cgw_groups	List available groups	—
cgw_list	List skills (all or by group); prepends the knowledge digest	group?
cgw_schema	Get parameter details for a skill (includes confirm flag)	skill
cgw_invoke	Invoke a skill (for skills with confirm=false)	skill, params?
cgw_invoke_confirmed	Invoke a skill that requires confirmation (native mode only)	skill, params?
cgw_recall	Search the knowledge base — facts, skill annotations, recipes	query?, kind?, tag?, limit?
cgw_remember	Persist a new knowledge note	kind, title, body, tags?, skillName?
cgw_update_note	Modify an existing note	id, plus any subset of fields
cgw_forget	Delete a note	id

Agent workflow

1. cgw_list                → sees org digest + skills grouped by purpose
2. cgw_recall(query=...)   → drills into details not covered by the digest
3. cgw_schema(skill=jira_issue) → sees parameters + confirm flag + which invoke to use
4. cgw_invoke(skill=jira_issue, params={key:"PROJ-1"}) → result (confirm=false)
   cgw_invoke_confirmed(skill=delete_issue, params={key:"PROJ-1"}) → result (confirm=true)
5. cgw_remember(...)       → saves a fact / annotation / recipe for the next session

Skill Configuration

Detailed guide: Creating Custom Skills (Русский)

Via extension UI

Extension settings (chrome://extensions → CorpGateway → Details → Extension options):

• Sidebar tabs: Skills / Recordings / Knowledge
• Skills mode: groups in the sidebar, skill list in the center, edit form in the right panel
• Recordings mode: named captures of your real portal usage that can be promoted into skills (see SKILLS.md)
• Knowledge mode: the agent's long-running memory — digest + recallable notes (see KNOWLEDGE.md)

Skill structure

Field	Description
Name	Function name for the agent (e.g. jira_issue)
Description	What the skill does (visible to agent)
URL	API endpoint with path parameters {id}
HTTP method	GET, POST, PUT, PATCH, DELETE
Origin URL	Where to execute the request from (if different from URL)
Body Template	JSON body template with {{param}} placeholders
Response Filter	Dot-notation paths for response filtering
HTTP headers	Custom headers (support {{param}})
Parameters	Name, type (String/Integer/Float/Boolean/Date), required, description

Parameter substitution

Location	Format	Example
URL path	{param}	/api/issues/{key} → /api/issues/PROJ-1
Body	{{param}}	{"text":"{{msg}}"} → {"text":"Hello"}
Headers	{{param}}	X-Data: {{token}} → X-Data: abc123
Query string	automatic	GET + remaining params → ?q=test&limit=10

Groups

Skills are organized into groups. Each group can be enabled/disabled — disabled groups and their skills are hidden from the agent.

Presets

Ready-made skill sets in presets/:

File	System	Skills
jira.json	Jira REST API	issue, search, comments, transitions
confluence.json	Confluence REST API	pages, search, spaces
gitlab.json	GitLab API	projects, issues, merge requests
mattermost.json	Mattermost API	channels, posts, users, search
outlook.json	Microsoft Graph API	mail, calendar, contacts

After import, replace URL placeholders (JIRA_URL, MATTERMOST_URL, etc.) with actual addresses.

Authorization

How the extension accesses APIs

1. Cookies: chrome.cookies API — the extension has access to cookies for all permitted domains
2. Authorization headers: chrome.webRequest.onSendHeaders — intercepts Authorization from all browser requests (Bearer tokens, JWT)
3. Fetch with credentials: requests are executed from the extension's Service Worker with credentials: 'include'

You don't need to enter passwords or API keys. If you're logged into a corporate system in Chrome — the extension automatically uses that session.

Security

┌─────────────────────────────────────────────────────────┐
│  Layer 1: HTTP API (cgw_mcp)                            │
│  • Bearer token on every agent request                  │
│  • Timing-safe token comparison                         │
│  • Rate limiting: 10 attempts/min per IP                │
│  • localhost only — not accessible from network         │
│  • CORS restricted to localhost                         │
│  • JSON-RPC method whitelist                            │
│  • Message size limit (1 MB)                            │
│                                                         │
│  Layer 2: WebSocket (cgw_mcp ↔ extension)               │
│  • Extension token for authentication                   │
│  • Mutual auth: HMAC challenge-response                 │
│  • Both sides prove knowledge of extensionToken         │
│                                                         │
│  Layer 3: Chrome Extension                              │
│  • Sender validation — web pages cannot send            │
│    commands to the extension                            │
│  • SSRF protection: private IP blocking, http(s) only   │
│  • Template validation: JSON/HTTP injection protection  │
│  • MCP connection — only via user button click          │
│  • Notifications on auth session expiry                 │
│                                                         │
│  Layer 4: Data                                          │
│  • Credentials not stored — uses Chrome session         │
│  • Tokens in cgw_mcp.json with 0600 permissions         │
│  • Skills in chrome.storage.local (encrypted by Chrome) │
│  • Audit log: last 100 invocations in session storage   │
│  • Tokens masked in logs                                │
│  • Confirmation modes: native (agent permissions) or OTP │
└─────────────────────────────────────────────────────────┘

Configuration

cgw_mcp.json

Location: ~/.corpgateway/cgw_mcp.json

{
  "port": 9877,
  "token": "agent-token",
  "extensionToken": "extension-ws-token",
  "mcpInstructions": "Instructions for the agent..."
}

Field	Description
port	HTTP port for cgw_mcp (default 9877)
token	Bearer token for the agent
extensionToken	Token for WebSocket connection from extension
mcpInstructions	Instruction text sent to agent on MCP initialize

Extension (Settings)

Field	Description
Instance name	Identification when using multiple browsers
MCP server URL	HTTP address of cgw_mcp
Extension token	extensionToken from cgw_mcp.json

Managing cgw_mcp

Install as daemon

Windows:

cd cgw_mcp
.\install.ps1              # install (Task Scheduler)
.\install.ps1 -Uninstall   # remove

Linux:

cd cgw_mcp
./install.sh               # install (systemd user service)
./install.sh uninstall      # remove

systemctl --user status cgw-mcp
systemctl --user restart cgw-mcp
journalctl --user -u cgw-mcp -f

macOS:

cd cgw_mcp
./install.sh               # install (LaunchAgent)
./install.sh uninstall      # remove

launchctl list | grep cgw-mcp
tail -f ~/.corpgateway/logs/cgw_mcp_launchd.log

Logs

Location: ~/.corpgateway/logs/

Rotation: last 7 days are kept. Format:

[2026-04-01T10:30:15.123Z] INFO cgw_mcp started on http://localhost:9877
[2026-04-01T10:30:20.456Z] INFO Extension connected: "Chrome Work"
[2026-04-01T10:31:05.789Z] INFO → Agent request id=1 method=tools/call

API Reference

POST /mcp

MCP JSON-RPC endpoint (Streamable HTTP). Requires Authorization: Bearer <token>.

GET /mcp

SSE keep-alive (MCP spec). Requires Authorization: Bearer <token>.

GET /health

{
  "status": "ok",
  "extension": true,
  "extensionName": "Chrome Work"
}

WS /extension/ws

WebSocket for the extension. Two-stage authentication:

1. Connect with ?token=<extensionToken>&name=<instanceName>
2. Mutual auth (HMAC challenge-response) — both sides prove knowledge of extensionToken

Multiple Browsers

If the extension is installed in multiple Chrome profiles:

1. Each profile has its own extension instance with its own skills
2. Only one can be connected to cgw_mcp at a time
3. The last one to connect displaces the previous
4. GET /health shows the name of the connected instance
5. Set different Instance names in settings for identification

File Structure

├── extension/                  # Chrome Extension (Manifest V3)
│   ├── manifest.json           # Permissions, service worker
│   ├── background.js           # WS connection to cgw_mcp, auth capture
│   ├── popup.html/js           # Connect button, groups
│   ├── options.html/js         # Full skill editor (3-column UI)
│   ├── lib/
│   │   ├── storage.js          # CRUD for skills/groups/recordings/knowledge
│   │   ├── executor.js         # Skill execution (fetch + substitution)
│   │   ├── recorder.js         # Traffic capture (filter, sanitize, persist)
│   │   └── mcp.js              # MCP JSON-RPC handler (meta-tools)
│   ├── _locales/               # i18n (en, ru)
│   └── icons/                  # Colored + gray icons
│
├── cgw_mcp/                    # MCP Server (Node.js)
│   ├── index.js                # HTTP + WebSocket server
│   ├── package.json
│   ├── install.sh              # Daemon installer (Linux/macOS)
│   └── install.ps1             # Daemon installer (Windows)
│
├── presets/                    # Ready-made skill sets
│   ├── jira.json
│   ├── confluence.json
│   ├── gitlab.json
│   ├── mattermost.json
│   └── outlook.json
│
├── docs/                       # Documentation
│   ├── README.ru.md            # README (Russian)
│   ├── SETUP.ru.md             # Setup guide (Russian)
│   ├── SKILLS.ru.md            # Creating custom skills (Russian)
│   ├── KNOWLEDGE.ru.md         # Knowledge layer (Russian)
│   └── SECURITY.ru.md          # Security architecture (Russian)
│
├── README.md                   # This file
├── SETUP.md                    # Installation guide
├── SKILLS.md                   # Creating custom skills
├── KNOWLEDGE.md                # Agent knowledge layer
└── SECURITY.md                 # Security architecture