ClawUI
Autonomous Orchestrator for CLI Coding Agents

Make your CLI coding agent observable, proactive, and autonomous.

CLI coding agents are powerful — but opaque. You launch a task, lose visibility, and hope for the best. ClawUI changes that. It wraps any CLI agent (Claude Code, OpenClaw, Codex, Pi) with a visual orchestration layer that decomposes complex goals into a Blueprint DAG, then drives execution through an Autopilot loop capable of self-reflection, adaptive re-planning, and error recovery — all while you watch, or walk away.

Autopilot paused after 20 iterations — awaiting human confirmation before release tasks.
The agent planned, executed, reflected, and adapted the blueprint autonomously.

---

Three Pillars

1. Observable

Every agent action is captured and visualized in real time. No more guessing what your agent is doing inside a terminal.

• Blueprint DAG — Complex tasks decomposed into a dependency graph of nodes, each with status, role, and execution history.
• Session Timeline — Every Claude Code interaction rendered as a structured, searchable timeline with I/O views.
• Autopilot Log — A decision-by-decision audit trail: what the agent chose, why, whether it succeeded, and how long it took.
• Blueprint Insights — Cross-cutting observations surfaced by the agent during execution (warnings, blockers, optimization opportunities).

2. Proactive with Reflection

The agent doesn't just follow a static plan — it thinks about what it's doing and adapts.

• Memory & Reflection — Every N iterations, the autopilot pauses to reflect: What's working? What patterns am I seeing? What should I try differently? Reflections are stored as per-blueprint memory and global strategy, injected into future decisions.
• Post-Completion Evaluation — After each node finishes, AI evaluates the result: COMPLETE, NEEDS_REFINEMENT (insert follow-up), or HAS_BLOCKER (create blocker sibling). The graph mutates itself at runtime.
• 19-Tool Decision Palette — The autopilot doesn't just "run the next node." It chooses from 19 actions: execute, enrich, split, reevaluate, skip, add nodes, rewire dependencies, triage suggestions, create insights, pause for human input, and more.
• Proactive Suggestions — When a session pauses, AI proposes the top next actions as one-click buttons.

3. Autonomous

Point it at a goal and let it drive — with guardrails.

• Autopilot Loop — An observe-decide-execute cycle that iterates over the blueprint, picking the highest-impact action each round. It handles the full lifecycle: planning, execution, evaluation, error recovery, and completion.
• Adaptive Re-Planning — When a node fails or reveals unexpected complexity, the agent can split it, add new nodes, rewire dependencies, or skip and move on — dynamically discovering new critical paths to reach the goal.
• Error Recovery — Failed nodes are resumed in the same session with targeted feedback. If resuming doesn't work, the agent tries splitting, re-enriching, or escalating. Multi-attempt tracking prevents infinite retry loops.
• Guardrails — Max iteration limits, idle-iteration detection with auto-pause, per-node attempt caps, and explicit pause-for-human-review when the agent encounters high-stakes decisions (destructive operations, ambiguous requirements, release gates).

---

Quick Start

Prerequisites

• Node.js 20+ and npm 10+
• Claude Code CLI (or another supported agent — see Multi-Agent Support)
• macOS, Linux, or Windows

Install & Run

# One command — installs on first run, then starts
npx @clawui/cli

# Or install globally
npm install -g @clawui/cli
claw-ui

Open the secure URL (with auth token) printed in the terminal. Custom ports: PORT=4001 FRONTEND_PORT=4000 npx @clawui/cli

From Source

git clone https://github.com/ccchow/ClawUI.git
cd ClawUI && npm install && npm run dev

First Blueprint

1. Click New Blueprint, enter your project directory and a goal (e.g., "add OAuth login with Google and GitHub providers").
2. AI analyzes your codebase and generates implementation-ready nodes with file paths, acceptance criteria, and dependency edges.
3. Click Autopilot to let the agent drive — or Run All for step-by-step manual execution.
4. Watch the autopilot log as the agent executes, evaluates, reflects, and adapts. Pause anytime to intervene.

Tip: A well-configured CLAUDE.md + MCP servers + skills dramatically improves node quality and reduces retries.

---

How Autopilot Works

         ┌──────────┐
    ┌───>│ OBSERVE  │ ── Collect node statuses, insights, suggestions, queue state
    │    └────┬─────┘
    │         v
    │    ┌──────────┐
    │    │ REFLECT  │ ── Every N iterations: update per-blueprint + global memory
    │    └────┬─────┘
    │         v
    │    ┌──────────┐
    │    │ DECIDE   │ ── Choose from 19 tools based on state + memory + strategy
    │    └────┬─────┘
    │         v
    │    ┌──────────┐
    │    │ EXECUTE  │ ── Run the chosen action (execute node, split, enrich, pause...)
    │    └────┬─────┘
    │         v
    │    ┌──────────┐
    └────│ EVALUATE │ ── Check result, update graph, loop or pause
         └──────────┘

The agent builds a state snapshot each iteration — a token-efficient summary of the entire blueprint — and reasons about the best next move. Its memory accumulates across iterations, so it gets smarter about the specific blueprint over time. When it finishes, it runs a global reflection to carry lessons into future blueprints.

---

Architecture

You (browser)  ◄──►  Next.js frontend (:3000)  ◄──►  Express backend (:3001)  ◄──►  Agent CLI
                                                              │
                                                        SQLite + JSONL
                                                      (100% local data)

• Four-layer data model: Raw JSONL (read-only) → SQLite index → Enrichments → App state. See Data Model.
• Pluggable agent runtimes: Claude Code, OpenClaw, Codex, Pi Mono. See Multi-Agent Support.
• Localhost-only: Both servers bind to 127.0.0.1. Remote access via tailscale serve.

---

Multi-Agent Support

Switch agent backend via the AGENT_TYPE environment variable:

Agent	AGENT_TYPE	Description
Claude Code	claude (default)	Anthropic's official CLI
OpenClaw	openclaw	Open-source agent framework
Pi Mono	pi	Lightweight coding agent
Codex CLI	codex	OpenAI's Codex CLI agent

AGENT_TYPE=openclaw npx @clawui/cli

---

Role System

Blueprints support a multi-role mechanism that tailors prompts and evaluation per node:

Role	Focus
SDE (Software Developer)	Implementation, code quality, architecture
QA (Quality Assurance)	Testing, validation, edge cases
PM (Product Manager)	Requirements, acceptance criteria, user stories

Set defaults at the blueprint level; override per-node as needed.

---

More Screenshots

Blueprint List	Task Node Detail

Session Timeline	Mobile View

---

Contributing

ClawUI can serve as its own development environment — point a Blueprint at this repo and let the agent build features through the very UI you're improving.

1. Fork & clone, run npm install && npm run dev
2. Create a Blueprint with your repo path as workspace
3. Describe what you want to build — ClawUI decomposes and executes it
4. Review & PR — see CONTRIBUTING.md

---

Disclaimer & Legal

ClawUI is an independent, unofficial community open-source project. It is not affiliated with, endorsed by, or associated with Anthropic PBC. "Claude" and "Claude Code" are trademarks of Anthropic.

ClawUI acts as a local GUI orchestrator and does not distribute, modify, or bundle proprietary CLI tools. Users must install and authenticate their chosen agent CLI independently. Designed exclusively for local, self-hosted usage by the authenticated individual.

License

MIT License (c) 2025-2026. See LICENSE for details.