Fleet · Steroids for AI coding

placeholder

v1.1.1 · MIT · works with any A2A agent

Steroids
for AI coding.

Your laptop, desktop, and EC2 — all running Claude Code, Codex, and Aider in parallel on the same repo. One command plans the work. Every machine takes a mission. GitHub is the only message bus.

Get started ▶ Watch how it works

$ npx fleetspark init copy

missions3

ships3

agents3

wall time23 min

How it works

Map-Reduce for software development.

Watch the full flow: your laptop becomes the commander, missions fan out through GitHub, ships work in parallel, branches converge. Scrub or step through to inspect any phase.

01 Idle 00:00 / 36s

Three machines. Zero coordination.

Your laptop is on Claude Code. Your desktop is idle. Your EC2 is still running from last sprint.

Live mission board

Watch the fleet from a single terminal.

fleet status — three ships, three branches, real-time heartbeats. Color-coded so a glance tells you who's working, who's stalled, who's ready to merge.

ship-laptop · M-001 · feature/auth · claude-code● HEALTHY

~/repo $ fleetspark ship --join git@github.com:you/repo.git

[claude-code] step 4/7 — adding session middleware…

✓ src/auth/oauth.ts (created)

✓ src/auth/session.ts (created)

[heartbeat] pushed fleet/state · +60s

▍ writing tests · 3 of 8…

ship-desktop · M-002 · feature/ratelimiter · codex● READY · PR #243

~/repo $ fleetspark ship --join git@github.com:you/repo.git

[codex] step 7/7 — opening PR…

✓ src/middleware/ratelimit.ts

✓ tests/ratelimit.test.ts (12 passed)

✓ CI green · all checks passed

→ PR #243 open, awaiting merge ✓

ship-ec2 · M-003 · feature/docs · aider● STALLED · 3 beats

~ $ fleetspark ship --join git@github.com:you/repo.git

[aider] step 2/5 — regenerating openapi.yaml…

✓ docs/api/auth.md (updated)

⚠ no heartbeat for 180s

[commander] shadow dispatch: re-queueing M-003 → ship-laptop

▍ waiting for next ship to claim…

drsti-dev-flow

Governed coding, not cowboy coding.

Most AI coding agents YOLO straight to a commit. FleetSpark's drsti-dev-flow plugin enforces three phases on every mission — spec, implementation, peer review — gated by maturity level. Risky changes get stricter scrutiny. Trivial ones breeze through. FleetSpark dogfoods this on itself.

Maturity-leveled gates

Auth, billing, schema changes — level 4 review. Doc tweaks — level 1. Configure triggers in .drsti/adapter.yml.

Three branches per mission

Spec, implementation, and review live on separate branches. Each phase is independently auditable.

One command, governed by default

fleet run --template drsti-dev-flow wires the whole flow. Or opt out and run lean.

$ fleet run --template drsti-dev-flow

Spark execution

The bits that turn three machines into one fleet.

Parallel dispatch, shadow re-runs when a ship goes dark, and fleet-wide context shared in seconds — not at the speed of "let me catch you up on Slack."

DISPATCH

DAG dispatch

The commander plans your goal as a dependency graph and pushes branches in topological order. Independent missions ship in parallel — automatically.

SHADOW

Shadow dispatch

A ship missed three heartbeats. Without waking you up, the commander reassigns the same mission to a healthy ship — first one to PR wins. No babysitting.

CONTEXT

Fleet brief

One mission generates context (an API schema, a migration plan). The brief is pushed once and every other ship picks it up before its next step. Map-reduce, but for context.

Agent ecosystem

Bring your own agent. Fleet doesn't care.

8 adapters ship out of the box. Mix Claude Code on your laptop with Codex on a desktop and Aider on a spot instance — all on the same plan, coordinated through GitHub.

🅒

Claude Code

@fleetspark/adapter-claude

shipped

Anthropic's CLI agent. The default for most fleet runs — battle-tested across hundreds of missions.

○

OpenAI Codex

@fleetspark/adapter-codex

shipped

OpenAI's terminal agent. Reasoning-heavy missions; runs alongside Claude Code on the same plan.

◇

Aider

@fleetspark/adapter-aider

shipped

The original local AI pair-programmer. Ideal for surgical edits and quick refactors on spot instances.

◆

OpenCode

@fleetspark/adapter-opencode

shipped

Open-source CLI agent. Run completely offline on your own hardware — no API keys required.

✦

Gemini CLI

@fleetspark/adapter-gemini

shipped

Google's Gemini CLI. Pair with Claude Code for multi-model redundancy across long-running runs.

⌥

Cursor

@fleetspark/adapter-cursor

shipped

Cursor in headless mode. Leverage its codebase-aware context engine inside fleet missions.

⚡

Amp

@fleetspark/adapter-amp

shipped

Sourcegraph Amp. Deep symbol navigation across large codebases — strong on refactor missions.

⌬

A2A / Custom

@fleetspark/adapter-a2a

shipped

Any A2A-compatible agent. Write a 50-line adapter and ship your own agent into the fleet.

Your adapter here

≈50 lines — see /adapters/

Why fleet

Every AI coding tool assumes one developer, one machine, one session.

Fleet is the only thing that turns the agents you already use into a coordinated swarm. The same Claude Code, Codex, and Aider — just multiplied.

Tool

Multi-machine

Multi-agent

Auto-planning

Auto-merge

Failover

Claude Code

Codex CLI

Aider

GitHub Copilot

⚡ fleetSpark

✓

✓ 8 agents

✓

See it in action

From zero to a coordinated AI coding fleet in four steps.

Initialize your project

One command sets up Fleet in any git repo.

Terminal

~/my-project $ npx fleetspark init

✓ Created .fleet/config.yml

✓ Created fleet/state branch

✓ Initialized empty FLEET.md

Fleet is ready. Run fleet command --plan to get started.

Plan your work

Describe what you want built. Fleet's LLM planner decomposes it into independent missions with dependencies.

Terminal — Commander

~/my-project $ npx fleetspark command --plan "Add OAuth login, fix rate limiter, update API docs"

⚡ Planning 3 goals...

✓ M-001 feat/oauth-login → branch created

✓ M-002 fix/rate-limiter → branch created

✓ M-003 docs/api-documentation → branch created, depends: M-001

⚡ Plan written to FLEET.md — 3 missions, 1 dependency

Commander monitoring started. Waiting for ships...

Launch ships on other machines

Each machine joins the fleet, picks up a mission, and starts its AI coding agent autonomously.

Machine 2 — Ship Alpha

~ $ npx fleetspark ship --join git@github.com:you/project.git

✓ Cloned repository

✓ Assigned mission M-001 (feat/oauth-login)

⚡ Starting claude-code adapter...

Heartbeat: pushing every 60s

● Agent working on M-001...

Machine 3 — Ship Beta

~ $ npx fleetspark ship --join git@github.com:you/project.git

✓ Cloned repository

✓ Assigned mission M-002 (fix/rate-limiter)

⚡ Starting gemini adapter...

Heartbeat: pushing every 60s

● Agent working on M-002...

Watch, merge, ship

Monitor from the terminal or your phone. Fleet auto-creates PRs, runs CI, and merges clean code.

Terminal — Status

~/my-project $ npx fleetspark status

⚡ Fleet Status Updated 5s ago

✓ M-001 feat/oauth-login ship-alpha completed → PR #12 merged

✓ M-002 fix/rate-limiter ship-beta completed → PR #13 merged

● M-003 docs/api-docs ship-alpha in-progress 2/4 steps

Ships: 2 alive │ Merged: 2/3 │ Time: 23 min

v1.1.1 · MIT · open source

Three machines.
Three missions.
Twenty-three minutes.

While you slept. While you were in a meeting. While you were doing literally anything else.

Start the fleet → View on GitHub

~/your-repo

$ npx fleetspark init

$ npx fleetspark command --plan \

"the thing you've been putting off"

↳ 7 missions planned. Open your laptop tomorrow.

Fleet · Steroids for AI coding

Steroids for AI coding.

Map-Reduce for software development.

Watch the fleet from a single terminal.

Governed coding, not cowboy coding.

Maturity-leveled gates

Three branches per mission

One command, governed by default

The bits that turn three machines into one fleet.

Bring your own agent. Fleet doesn't care.

Every AI coding tool assumes one developer, one machine, one session.

See it in action

Initialize your project

Plan your work

Launch ships on other machines

Watch, merge, ship

Three machines.Three missions.Twenty-three minutes.

Steroids
for AI coding.

Three machines.
Three missions.
Twenty-three minutes.