Captain Claw – Local AI Agent with Multi-Agent Orchestration

TL;DR

TL;DR: Captain Claw is an open-source local AI agent that orchestrates multi-agent workflows via Basna ensemble routing and an Agent Forge archetype library, ships with 47 built-in tools, and runs fully offline on your own hardware.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: github.com/kstevica/captain-claw
• License: MIT (verified via LICENSE file in repository)
• Latest release: v0.6.0 (verified via GitHub Releases API, published 2026-06-16)
• HN launch thread: news.ycombinator.com/item?id=47261534
• Source last checked: 2026-06-18 (commit main, README + release notes v0.6.0)

What Is Captain Claw?

Captain Claw is an open-source AI agent framework written in Python (3.11+) that runs entirely on your local machine. It wraps a pool of LLM providers — OpenAI, Anthropic Claude, Google Gemini, Ollama, and OpenRouter — behind a single orchestration layer, and exposes a web-based UI called Flight Deck alongside a terminal interface.

The core innovation in recent releases is Basna (v0.5.7+), a network-source ensemble system where a router selects the smallest set of specialist archetypes, spawns them in parallel, and merges their outputs weighted by learned reliability. It self-improves over time based on your feedback.

Setup Workflow

Step 1: Install

pip install captain-claw

Or use the standalone binary (no Python required):

# After installing, entry points are:
captain-claw          # full CLI
captain-claw-fd       # Flight Deck web UI
captain-claw-agent    # agent-only mode

Step 2: Configure API Keys

Set environment variables for your LLM providers:

export OPENAI_API_KEY="sk-....port ANTHROPIC_API_KEY="sk-a...port OPENROUTER_API_KEY="sk-o...nFor local Ollama:

```bash
export OLLAMA_BASE_URL="http://localhost:11434"

Step 3: Launch Flight Deck

captain-claw-fd

Open http://localhost:8000 to access the web dashboard.

Step 4: Spawn Your First Agent

From the Flight Deck UI, click Spawn Agent, pick a model tier (Reasoning / Balanced / Fast / Long context), and choose from the archetype gallery. Captain Claw ships with approximately 20 curated archetypes including Deep Researcher, Code Reviewer, Project Coordinator, and Deal Screener.

Deeper Analysis

Basna Ensemble Routing

Basna is the headline feature of v0.5.7 and later. Unlike a fixed council of agents, Basna uses a router that dynamically selects which archetypes to invoke based on the task. It runs them in parallel, blind, and merges outputs weighted by each archetype’s learned reliability on your specific feedback patterns.

task → router selects archetypes → parallel spawn → weighted merge → final output

The system learns per-archetype reliability. A thumbs-up or thumbs-down on any agent output adjusts its weight for future routing. This means Basna gets better the more you use it.

Agent Forge and Archetypes

Agent Forge is the team-composition interface. An archetype is a curated, reusable agent profile with a system prompt, default model tier, and task scope. The base library ships roughly 20 archetypes; you can extend it on the Library page — either by hand or generated from a prompt. Overriding a base archetype (same ID) or adding new ones is fully supported.

Flow Automation

Since v0.5.0, Captain Claw includes a Flow language — a deterministic automation DSL that composes agents, tools, and control structures (loops, branches, parallel spawns, error handlers) into shareable programs. Flows can be authored visually in Flight Deck or written as code in the built-in DSL editor.

Key Flow primitives:

• gosub <flow> — call a sub-flow as a function, pass args, use return values
• spawn <flow> — launch in background, collect with join
• foreach <var> in <list> — iterate in parallel (spawn) or sequential (gosub)
• wait until contains "..." — pause flow until an inbound message arrives
• error handler steps for resilient failure handling

Built-in Tools

The README advertises 47 built-in tools covering file operations, web search, image vision, code execution, and integrations. Specific tools mentioned include:

• web_fetch and web_fetch_batch — parallel web scraping with automatic headless browser fallback for JS-rendered pages
• file_* operations — read, write, edit, search
• synthesize_flow — AI-compiles a plain-English description into a runnable Flow
• WhatsApp and Meta Ray-Ban glasses integrations (channel-agnostic — same controls work everywhere)
• MCP server support with a live Connections panel

Multi-Provider Support

The model tier system (v0.5.6+) lets you configure separate provider/model/key/base-URL settings per tier (Reasoning / Balanced / Fast / Long context), persisted per user. Agents resolve to a concrete model at spawn from these saved tiers.

Practical Evaluation Checklist

• Runs fully offline (Ollama + local models) — no cloud dependency required
• Multi-agent orchestration with learned routing (Basna) — adaptive, not fixed council
• Visual Flow builder with DSL editor — deterministic automation without YAML hell
• Archetype library extensible from within the UI — no config-file editing needed
• Connections panel monitors MCP server health — avoids silent integration failures
• Token meters shown live per turn — no surprises on billing with paid providers
• Desktop standalone available — Docker-free local process mode with reliable file writes

Security Notes

• All agent reasoning and tool calls stay on your machine — no telemetry to third parties (verified from README, no external analytics mentioned)
• MCP server connections are read-only probes with a 10-minute polling interval
• API keys are sourced from environment variables, never hardcoded
• Basna sessions and run logs are stored locally in the app’s data directory

FAQ

Q: Does it work without an internet connection? A: Yes. With Ollama configured as a provider, Captain Claw runs entirely offline. Cloud providers (OpenAI, Anthropic, etc.) require internet, but the architecture does not force you to use them.

Q: How does Basna differ from a fixed agent council? A: A council runs the same agents every time. Basna’s router dynamically selects which archetypes to invoke per task, runs them in parallel, and learns which ones perform best for your specific feedback patterns. The routing improves over time.

Q: Can I run this on macOS? A: Yes. The project is cross-platform Python. The desktop standalone binary also supports macOS. Homebrew or pip installation works on macOS directly.

Q: How many agents can it run simultaneously? A: The practical limit depends on your hardware (VRAM for local models, or API rate limits for cloud providers). The architecture is designed for multi-agent pools; there is no hard-coded cap mentioned in the docs.

Q: Is there a mobile interface? A: No dedicated mobile app, but the web UI is responsive and WhatsApp integration is built-in, giving you chat-based agent control from any phone.

Conclusion

Captain Claw is a serious local AI agent framework that goes beyond single-agent chatbots. The Basna ensemble routing, archetype library, Flow automation engine, and extensible tool set make it one of the most capable self-hosted agent platforms available. Version 0.6.0 is the current release with a published MIT license.