UniClaw 🦞

A personal AI agent that routes tasks from any browser-based chat (Slack, Discord, Teams, etc.) to terminal-based AI agents (Claude Code, Opencode, Aider, etc.).

Slack is the current implementation, but the Chrome DevTools approach works with any web-based messaging platform.

Why UniClaw?

Similar to OpenClaw but optimized for developer workflows with universal agent support.

Feature Comparison

Feature	OpenClaw	UniClaw
Agent	API-based LLMs (OpenAI, etc.)	Any terminal agent (Claude Code, Opencode, Aider)
Channels	20+ native integrations	Any browser-based chat (Slack, Teams, Discord, etc.)
Tools	bash, read/write, browser	Full agent capabilities
Setup	npm + wizard	Node + Chrome debug
Complexity	Full platform (~265K LOC)	Lightweight (~250 LOC)
Cost	API usage fees	Your agent's cost model

Key Advantages Over OpenClaw

1. Multi-Agent Parallel Execution

OpenClaw: Single agent per session. If busy → request is queued or blocked.

UniClaw: Multiple independent agents. One can work while another responds.

# jarvis running Long task
jarvis: run full eZe test suite (takes 30 mins)
# friday can observe, query status,
or do other work
friday: what's the status of jarvis's tests? 
friday: check jarvis's logs for errors 
friday: meanwhile, review this PR

2. No Blocking on Long Tasks

Each pane is an independent agent session
Long-running tasks don't block other queries
Monitor progress from your phone while tests run

3. Cross-Agent Observability

One agent can check another's output files
Real-time log streaming to Slack
Debug one task while another runs

4. Universal Agent Support

Claude Code, Opencode, Aider, Codex CLI, Gemini CLI
Any terminal-based AI agent works
Mix different agents across panes

5. Leverage Existing Agent Ecosystem

OpenClaw: Creates new agents from scratch via API — no skills, no memory, no context.

UniClaw: Uses your existing Claude Code sessions with full ecosystem:

Skills: All /skill commands work (git, bash, code analysis, etc.)
Memory: Claude remembers your codebase, preferences, patterns
Context: Project-specific knowledge from .claude/ directory
No retraining: Your agent already knows your project

6. Full System Access

OpenClaw: Limited to defined tool APIs
UniClaw: Full agent capabilities (git, tests, file system, terminal)

Use Case Examples

Scenario	OpenClaw	UniClaw
Run 30-min test suite	Blocked until done	Pane 1 runs, Pane 2 free
Check status mid-task	/status shows queue	Ask Pane 2 to check Pane 1
Parallel code reviews	Queue requests	Each pane reviews different PR
Debug while building	Wait for build	Build on Pane 1, debug on Pane 2

Architecture

Concept Mapping (OpenClaw → UniClaw)

OpenClaw Concept	UniClaw Equivalent	Description
Gateway	`uniclaw-agent.js`	Central control plane, routes messages
Channel	Chrome Web	Chrome + Slack Web
Agent/Sessions	iTerm Pane + Claude Code	Terminal-based AI agent
Tools	Agent's native tools	Full system access (git, fs, terminal)

System Diagram

Data Flow

Features

Multi-agent routing: jarvis: → Pane 1, friday: → Pane 2, eleven: → Pane 3 (expandable)
Universal agent support: Claude Code, Opencode, Aider, any terminal agent
Async observation: Watch logs from phone while tasks run
Chunked output: Long results split into multiple Slack messages (~3500 chars each)
Duplicate detection: Hash tracking prevents re-processing
Response detection: Skips commands that already have <agent>: responses
No blocking: Parallel task execution across panes

Behavior details:

getLatestMessageForAgent() only returns the most recent command (iterates from end, breaks on first match)
When busy (pending_task <agent> exists), the new command's hash is NOT saved
After task completes, only the latest command is checked - any commands sent in between are skipped

Workaround: Wait for <agent>: confirmation before sending the next command, or use a different agent (e.g., friday: instead of jarvis:).

Files

File	Purpose
`uniclaw-agent.js`	Monitors Slack DM, dispatches tasks, sends results
`notify-other-session.sh`	Sends input to iTerm pane via AppleScript
`run.sh`	Start/stop/status management script
`start-headless.sh`	Start headless Chrome with debug port for cookie export
`start-chrome-debug.sh`	Start Chrome with Slack session

Prerequisites

Node. js with ws module
```
cd ~/uniclaw && npm install ws
```
iTerm2 (for split panes)
Chrome (for initial Slack login only)

Quick Start

First-Time Setup

# 1. Start debug Chrome (uses separate profile, won't affect your regular Chrome)
./start-chrome-debug.sh
# 2. In the new Chrome window:
# - Go to https://app.slack. com
# - Login to Slack
# - Open your DM (messages to yourself)
# 3. Open iTerm and create split panes (Cd+D)
# - Pane 1: for jarvis
# - Pane 2: for friday
# 4. Start Claude Code in each pane
# Pane 1: claude
# Pane 2: claude
# 5. Start UniClaw
./run.sh start

Daily Startup (after first-time setup)

# 1. Start debug Chrome (reuses saved Slack session)
./start-chrome-debug.sh
# 2. Start Claude in iTerm panes
# Pane 1: claude
# Pane 2: claude
# 3. Start UniClaw
./run.sh start

That's it! Send messages in Slack DM Like 'jarvis: hello to test.

How It Works

Chrome Instance	Port	Purpose
Debug Chrome	9222	Export Slack cookies (visible, for login)
Headless Chrome	9223	Monitor Slack messages (invisible, runs in background)

On ./run.sh start:

Cookies are exported from Chrome debug → headless Chrome
Headless Chrome opens Slack with your session
Agent polls headless Chrome for new messages
UniClaw determines which iTerm pane receives the task
UniClaw prefixes messages for these prefixes
Responses are tagged with the agent name (e.g., jarvis:)

Usage

Agent Names

UniClaw uses Agent Names as message prefixes to route tasks to the correct terminal pane:

Agent Name	Target Pane	Example
`jarvis:`	Pane 1	`jarvis: run tests`
`friday:`	Pane 2	`friday: check git status`
`eleven:`	Pane 3	`eleven: review this PR`

How It Works:**

You prefix your Slack message with an Agent Name (e.g.g., jarvis:)
UniClaw determines which iTerm pane receives the task
Responses are tagged with the agent name (e.g., jarvis>)

Examples

Basic commands

jarvis: pwd friday: check git status

Long-running tasks

jarvis: run all e2e tests

Query status while jarvis is busy

friday: what's in jarvis's result file? friday: tail the test logs

Parallel work

jarvis: refactor the auth module friday: write tests for the API eleven: review the PR

Dynamic Agent Spawning

Agents can spawn new Claude Code sessions for other agents. This is useful when you need more parallel workers or when an agent is offline.

# Ask jarvis to bring eleven online
jarvis: eleven is offline, start a new claude code session in pane 3 for eleven

# jarvis will:
# 1. Open a new iTerm pane or select pane 3
# 2. Run `claude` command to start Claude Code
# 3. eleven is now ready to receive tasks

Now you can use eleven:
eleven: review PR 456

Use cases:

Scale up on demand: Start with one agent, spawn more when workload increases
Recovery: If an agent crashes, ask another agent to restart it
Specialized agents: Spawn agents in different directories for different projects

Example: Scale up for a big task

jarvis: I need help. Start friday in pane 2 with working dir ~/project-b

# jarvis spawns friday, then you can parallelize
jarvis: refactor module A
friday: refactor module B

Note: The spawning agent needs appropriate permissions to run AppleScript and control iTerm.

Management

./run.sh start       # Start agent
./run.sh stop        # Stop agent
./run.sh stop-chrome # Stop Chrome 9223
./run.sh status      # Check if running
./run.sh restart     # Restart agent (keeps Chrome running)
./run.sh logs        # Follow logs

Note: restart only restarts the agent, not Chrome. Use stop-chrome separately if needed. Message Flow

uniclaw-agent.js polls Slack DM every 10 seconds via headless Chrome
Detects messages starting with agent names (e.g., jarvis:, friday:)
Checks if response already exists (e.g. jarvis:) - skip if so
Writes task to pending_task_ (e.g., pending_task_jarvis)
Calls notify-other-session.sh which:

Selects the target iTerm pane
Writes the task prompt
Presses Enter to submit

Terminal agent receives the task and executes it
Agent writes results to task_result_ (in chunks if long)
uniclaw-agent.js picks up result and sends to Slack DM

Configuration

Edit the CONFIG section in uniclaw-agent.js:

const CONFIG = {
  // Chrome DevTools Protocol
  cdpPort: 9223,        // Headless Chrome port
  pageId: null,         // Auto-detected on startup
  checkInterval: 10000, // 10 seconds

  // Working directory for task files
  workDir: '/Users/yourname/uniclaw',

  // Target DM username - always send messages to this DM
  // Set to null to send to current active chat (old behavior)
  dmTarget: 'Your Name [Team]you',

  // Agent Name to iTerm Pane mapping
  agents: {
    'jarvis': 1,
    'friday': 2,
    'eleven': 3,
  }
}

DM Target

The dmTarget option ensures all responses are sent to a specific DM, even if you're viewing a different channel in Chrome:

Set to your Slack display name as shown in sidebar (e.g., 'Your Name [Team]you')
Set to null to send to whatever chat is currently open (old behavior)

Add more agents

Add new entries to CONFIG.agents:

agents: {
  'jarvis': 1,
  'friday': 2,
  'eleven': 3,
  'ultron': 4,  // Any string works
}
## Extending
### Support other chat platforms
UniClaw's Chrome DevTools approach works with **any browser-based chat platform**:
- **Discord**: Update DOM selectors for Discord web
- **Microsoft Teams**: Change selectors for Teams web
- **Lark**: Adapt for Lark web interface
- **DingTalk**: Support DingTalk web

Just change the CSS selectors in `uniclaw-agent.js` — no API integrations, no bot setup, no webhook configuration needed.



## Troubleshooting
### Debug Chrome not starting
```bash
# Check if port 9222 is in use
curl -s http://127.0.0.1:9222/json/version
# If nothing responds, start debug Chrome
/start-chrome-debug.sh

Headless Chrome not connecting to Slack

# Check headless Chrome status
curl -s http://127.0.0.1:9223/json/version
# Force restart headless Chrome with fresh cookies
./run.sh restart

Agent not detecting messages

# Check agent logs
./run.sh logs
# Verify Slack page is loaded in headless Chrome
curl -s http://127.0.0.1:9223/json | grep -i slack

Slack session expired

# 1. Open debug Chrome and re-login to Slack
./start-chrome-debug.sh
# (login to Slack in the browser)
# 2. Restart UniClaw to export fresh cookies
./run.sh restart

Messages not being picked up （stale page）

# Restart to get fresh headless Chrome session
./run.sh restart

Check agent logs

./run.sh logs
# or
tail -f agent.log

Duplicate task detection

Delete 'last_hash_* files (e.g., last_hash_jarvis') to reset.

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 15 Commits
docs		docs
.gitignore		.gitignore
README.md		README.md
architecture.svg		architecture.svg
config.sh		config.sh
datapath.svg		datapath.svg
export-cookies.js		export-cookies.js
import-cookies.js		import-cookies.js
notify-iterm-api.py		notify-iterm-api.py
notify-other-session.sh		notify-other-session.sh
package-lock.json		package-lock.json
package.json		package.json
run.sh		run.sh
start-chrome-debug.sh		start-chrome-debug.sh
start-headless.sh		start-headless.sh
sync-cookies.js		sync-cookies.js
uniclaw-agent.js		uniclaw-agent.js
wechat-article.md		wechat-article.md

Folders and files

Latest commit

History

Repository files navigation

UniClaw 🦞

Why UniClaw?

Feature Comparison

Key Advantages Over OpenClaw

1. Multi-Agent Parallel Execution

2. No Blocking on Long Tasks

3. Cross-Agent Observability

4. Universal Agent Support

5. Leverage Existing Agent Ecosystem

6. Full System Access

Use Case Examples

Architecture

Concept Mapping (OpenClaw → UniClaw)

System Diagram

Data Flow

Features

Files

Prerequisites

Quick Start

First-Time Setup

Daily Startup (after first-time setup)

How It Works

Usage

Agent Names

How It Works:**

Examples

Basic commands

Long-running tasks

Query status while jarvis is busy

Parallel work

Dynamic Agent Spawning

Example: Scale up for a big task

Management

Configuration

DM Target

Add more agents

Headless Chrome not connecting to Slack

Agent not detecting messages

Slack session expired

Messages not being picked up （stale page）

Check agent logs

Duplicate task detection

License

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages