Most AI tools give you one agent in a chat box. Captain Claw Flight Deck is a multi-agent command center — spawn specialist teams, run six orchestration modes, ship software with the Code pipeline, and compose deterministic Flows. All on your own machine.
Agent Desktop — spawn a team of specialists, monitor them live, and chat with any one from a single dashboard
6
Orchestration modes
47
Built-in tools per agent
∞
Agents per fleet
6
Memory layers shared
Flight Deck is the cockpit for everything that follows. Spawn agents, watch them work in real time, chat with any one of them, and coordinate the whole fleet — without leaving the page.
Every agent is a card — its model, persona, current task, and status. Open any one to chat, inspect files, or read its logs without pausing the others.
Mix GPT, Claude, Gemini, DeepSeek, and Ollama across one fleet. Give each agent only the tools its role needs — from 48 built in.
Per-agent token and cost meters, plus a full activity log with traceable lineage for every tool call, consultation, and hand-off.
The fastest way in — pick an archetype and start talking immediately. The agent spawns hidden from the desktop with the full chat experience (plan mode, attachments, next steps); hit Promote to desktop when you want it on the canvas.
Built-in Flight Deck modules
One agent isn't always the answer — and one orchestration strategy never is. Captain Claw ships six distinct modes, from ensemble reasoning to a full engineering pipeline — each tuned to a different kind of work.
Every mode draws from a library of 31 built-in archetypes — each a role with a tuned prompt, model, and tool set. Pick them by hand, let Agent Forge assemble a team, or fork one into your own.
The Archetypes library — 31 specialists, each editable, spanning research, writing, engineering, data, ops, finance, and multimedia.
Deep Researcher · Market & Competitor Scanner · Fact Checker
Editor & Long-form Writer · Outbound Communications · Social Repurposer
Software Implementer · Code Reviewer · Software Architect · Refactor & Simplifier · Debugger
Data Analyst · Report Builder
Project Coordinator · Monitor & Watchdog · Triage & Router · Concierge · Inbox & Calendar Manager
Deal Screener · IC Memo Preparer · Portfolio Watcher · Visual Document Extractor · UI & Screenshot Reviewer · Brand & Visual Design Reviewer
Describe the objective — "scout startups in Croatia, research the founders, build an investment pipeline" — and an LLM designs the whole org: lead coordinator, specialist roles, the right model and tool set per agent, and Standard Operating Procedures that keep them on rails.
Describe the objective
Plain English — no role-by-role spec required
AI designs the team
Roles, models, tools, and fleet-wide SOPs — all generated and editable
Review, customize, spawn
Tweak anything, then deploy the entire team in one click
Convene a financial analyst, a tech strategist, and a creative director. Each agent self-scores its suitability per topic, then chooses how to engage — answer, challenge, refine, or broaden — across moderated rounds. The moderator synthesizes conclusions and every agent votes. Export the minutes as markdown.
2-N
Agents per council
8
Session types
5
Agent actions
MD
Export minutes
Where Council debates and Forge staffs, Basna and Vatra are about quality of output — through independence and through collaboration.
N agents tackle the same question independently — no peeking at each other's work. Their answers are merged by reliability into one high-confidence result. Diversity beats a single point of failure.
Agents collaborate on a shared blackboard, each owning sections of the deliverable, reading what others wrote, and improving it over review rounds. Ideal for long documents that need many hands and one voice.
Describe what you want built. A router sizes the job — small edits go straight to a specialist; real features run a plan-gated, independently-reviewed pipeline, every phase committed to a real git repo.
↺ A capped fix loop (≤3 rounds) re-reviews only the delta — style nits never trigger a fix round.
A code reviewer, a security reviewer (CVSS-ranked), and a QA engineer that actually runs your test suite — none of them wrote the code they judge.
A per-repo symbol map in SQLite + FTS5. Agents query pointers, not source dumps, and stop re-reading your tree. Git-hash gated, so freshness costs pennies.
Plan, build, each review and fix land as commits in the folder's own repo. Any commit opens a colorized diff and is one click from rollback.
Point Code at an existing local repo (read-write or read-only). Run each role on its own model — DeepSeek to type, a reasoning model to plan, or 100% local Ollama.
Flows are not prompt chains. They're real programs with a deterministic spine — triggering, routing, sequencing, guardrails — where AI judgment only fires for the steps that truly need it.
Every Flow is a small text program, validated by a deterministic parser — not interpreted by an LLM reading YAML. The model is never the source of truth; the parser is.
Compose — flows call flows with gosub, pass arguments, and return values. Flows are functions.
Parallelize — spawn / join run work concurrently. Three lookups that took 9s in series finish in 3.
Recover — on error handlers and inline status branches. No silent half-failures.
Control mid-flight — /flow status, pause, stop from any channel, or buttons in the run log.
flow "Competitor Watch"
trigger cron "0 9 * * 1"
step fetch:
spawn pricing, features, news
join -> results
step diff:
agent
prompt: "Flag changes vs {{memory.last}}"
step report:
gosub "Build Report" with {{steps.diff.output}}
on error -> notify
output -> email, slack
Write the DSL, build it visually, or describe it in plain English — all round-trip losslessly.
The same flow, two views — drag it together in the visual builder or edit the DSL directly. They round-trip losslessly through one parser.
The synthesize_flow tool turns a repeated task into a validated, call-only flow stored in a
curated scratch space. A synthesized flow earns its way to permanence — three clean runs make it a ⭐ candidate,
three failures quarantine it. Promote the good ones into your permanent flows. Agent-written automations can never borrow your vetted
authority until you approve them.
Two modes that trade speed for correctness — structured execution and frontier-quality reasoning.
Turns a single request into a reviewable DAG of 3-8 steps, runs them under the orchestrator, judges each against acceptance criteria, and automatically re-plans whatever fails.
Decompose → review → execute → verify → re-plan
Inspect and edit the DAG before anything runs
Failed steps re-plan instead of cascading errors
Frontier-quality answers through multiple rollouts, self-consistency voting, and diverse-lens critics that attack the draft from different angles before it's returned.
Multiple independent rollouts per question
Self-consistency voting across candidates
Diverse-lens critics catch what redundancy can't
Every fleet runs behind input, output, and tool-execution guards with approval gates for sensitive operations — and a full audit trail with traceable lineage across all agents.
Orchestration is only as good as what the agents remember. Every mode runs on the same six-layer memory stack — so context, insights, and decisions carry across sessions and across agents.
Working Memory
Current conversation context with smart compaction
Semantic Memory
Vector + BM25 hybrid search across all sessions
Deep Memory
Long-term archive with full-text search
Insights
Auto-extracted facts, contacts, decisions, and deadlines
Nervous System
Autonomous dreaming that discovers non-obvious connections
Self-Reflection
Periodic self-assessment that generates improvement directives
1 · Working Memory
the live conversation
2 · Semantic Memory
vector + BM25 recall
3 · Deep Memory
full-text long-term archive
4 · Insights
facts, decisions, deadlines
5 · Nervous System
autonomous dreaming
6 · Self-Reflection
improvement directives
From the live turn down to autonomous self-review — every layer feeds every agent.
Every one of your agents reads and writes into a shared Virtual File System — a sandboxed workspace, scoped per Flight Deck user, that lives apart from your host machine. When one agent produces a report, a data extract, or a generated asset, it lands in the VFS where any of your other agents can pick it up. No copy-pasting between agents, no scattered downloads.
Shared across your fleet — all of a user's agents and sessions read and write the same tree.
Browsable in Flight Deck — explore, preview, and download files from the Agent Folders view.
Host-sandboxed, per user — each Flight Deck user gets their own isolated tree; agent file ops never touch the host disk.
The handoff layer — backs inter-agent file transfer and the shared clipboard across the crew.
/vfs
├─ reports/
├─ competitor-watch.html · research-agent
└─ q2-summary.pdf · report-builder
├─ data/
└─ contacts.sqlite · data-analyst
└─ assets/
└─ cover.png · design-reviewer
One tree, written by many agents — every artifact attributable to the agent that made it.
Between conversations, Captain Claw doesn't go idle. Its nervous system keeps running — dreaming over everything it knows, forming intuitions, tracking the intentions you've set, and watching the whole fleet for patterns. The Observatory is the live window into that inner life.
Notes-to-self it keeps revisiting — "watch whether this claim holds," "track that agent's behavior" — until they resolve.
A timestamped feed of thoughts and dream-cycle discoveries, each tagged with what it touched and how confident it is.
Contradictions are tracked openly rather than papered over — intuitions mature across dream cycles before they surface.
It watches every agent at once — repeated failures, idle signals, shared blockers — and flags what needs your attention.
When it notices something worth acting on, it can reach out — a nudge over WhatsApp, a scheduled task — but only with your approval.
Every thought traces back to a source — the six-layer memory and full lineage make the inner monologue auditable.
Self-hosted, fully private, no vendor lock-in
One-command deployment with Docker Compose. Persistent data, auto-restart, production ready.
Install from PyPI in seconds. Python 3.11+ with interactive onboarding to configure models.
Pre-built executables for macOS, Linux, and Windows. No Python required — download and run.
$ pip install captain-claw
$ flight-deck
Flight Deck running at http://localhost:25080
A full, live Flight Deck in your browser — no sign-up, no install. Drive a real fleet of agents in minutes.
This isn't a video or a stripped-down sandbox — it's the full Flight Deck with all 48 tools and every orchestration mode. Spawn agents, watch the nervous system think, and ship a Flow. Bring your own key and go.
Open the live Flight DeckIn 5 minutes, try…
› Forge a team
Describe a goal → AI staffs it
› Run a Basna ensemble
Blind specialists, merged answer
› Convene a Council
Agents debate, then vote
› Author a Flow
Visual builder or DSL
› Watch the Observatory
The fleet's stream of thought
› Switch models mid-chat
GPT ↔ Claude ↔ Gemini
Prefer to keep it permanent and private? Install Captain Claw on your own machine — pip install captain-claw.
Captain Claw is MIT licensed. Your data stays on your machine. Bring your own API keys, or run 100% locally with Ollama.
Built solo, growing fast. Every ⭐ helps another builder discover Captain Claw — if it looks useful, would you give it a star?
Star on GitHub