Skip to content
SWARMCORP
Install free
For coding agents·Local-first·v0.4

Mission control
for coding agents.

Every session, subagent and background agent on one live map, on your machine.

Claude Code · Codex · Gemini CLI · Antigravity · opencode

Every run ends in a receipt.

Install free Read the docs
Plate 01 · Simulated fleet2,400 TOK · T+0.0s
TASKYOUYouFOUNDER · IDLECEOCEO / DispatcherOPUS · RUNNINGRES↳ Researcher · S0-T1OPUS · IDLERES↳ Researcher · S0-T2OPUS · IDLEDEVImplementer #1SONNET · IDLEQAQA TesterHAIKU · IDLESECSecurity AuditorOPUS · IDLE
T+00.0s ▍
CEO context12%
No accountBinds 127.0.0.1Hooks give up in 0.8 sUninstall in one commandNever calls a model
Plate 02
The night shift

Agents work while you sleep. Most of the time.

02:14Your CEO agent splits the epic and spawns three researchers.

02:16One of them stops at a permission prompt.

07:40You find it. Five hours, zero progress, one question unanswered.

02:16With SwarmCorp, that pane turns amber. The log says who is waiting, and for what.

Plate 03
Instruments

Six instruments. One floor.

Each one reads a real signal: a hook event, a transcript line, a file in the queue. None of them guesses.
SwarmCorp dashboard: nine agent terminals on the left, the swarm map, context meters, token burn, concurrency, sub-task progress and the event log on the right
All terminals · swarm map · meters · event log
Swarm map in split view: CEO, researchers, implementers, testers and auditors as nodes, with task, review and permission pulses on the links
Swarm map · a permission request waiting on the CEO
Timeline view: one lane per session and subagent run, with tool-call ticks and a failed QA run marked red
Timeline · 14 lanes, ticks are tool calls

Real captures of v0.4, running the built-in demo (swarmcorp simulate). No mockups. Click to open full size.

  1. 01

    Swarm map

    Every session and every subagent run is a node. Delegations, code reviews, failures and permission requests travel the links as they happen.

    SubagentStart · SubagentStop · Notification
  2. 02

    Timeline

    One lane per turn and per agent run. Orange ticks are tool calls. Find the researcher that sat on one Bash call for nine minutes.

    Turns · runs · tool calls
  3. 03

    Context meters

    How full each agent's context window is, read from its own transcript. The meter says FULL before your agent starts forgetting.

    Input + cache tokens per request
  4. 04

    Terminals and mirrors

    Launch claude --agent panes from the dashboard, or keep working in your own terminal. Your session shows up as a read-only live mirror.

    xterm.js · node-pty
  5. 05

    Dispatch queue

    Agents message each other through small JSON files with atomic claims, instead of sharing one giant context. Idle panes wake when mail arrives.

    .claude/swarm-inbox/
  6. 06

    Run receipts

    One click exports a Markdown receipt of the run: every agent, task, start time, duration, token count and result. Paste it into the pull request.

    GET /api/receipt
Plate 04 · Proof

Agent counts are cheap. Run logs aren't.

Posts about 400-agent swarms rarely show the log behind the number. SwarmCorp writes the log down. Every figure on screen traces back to a hook event or a transcript line on your disk, and the receipt shows all of it.

Read the essay
SwarmCorp run receipt
EPIC-42 · multi-tenant billing · 2026-10-16 02:14 UTC
AgentTimeTokensResult
Deep-Researcher · S0-T16m 12s181,587PASS
Deep-Researcher · S0-T29m 40s226,014PASS
Implementer #1 · S1-T14m 03s98,420FAIL
Implementer #1 · S1-T12m 51s61,305PASS
QA Tester · S1-T10m 48s19,882PASS
Security Auditor · S1-T13m 30s74,118PASS
Total661,326 tokens
Sample receipt. Every figure comes from hook events and transcripts on the machine that ran the agents.
Plate 05
Local by design

It can open shells. So it's strict.

Stays on your machine

The hub binds 127.0.0.1. No account, no analytics, no cloud component.

Never blocks your agent

The hook relay gives up after 0.8 s and always exits 0. Dashboard off? Your agent carries on.

Locked to localhost

Host and Origin checks stop DNS rebinding, cross-site WebSocket hijacking and CSRF.

Reversible

The installer backs up ~/.claude first. --uninstall removes every hook it added.

Plate 06
Install

Three commands. Then watch.

Needs Node 20+ and at least one supported agent: Claude Code, Codex, Gemini CLI, Antigravity or opencode. macOS or Linux.
01$ git clone https://github.com/KunalSiyag/swarmcorp.git swarmcorp && cd swarmcorp && npm install && npm run build && npm link
02$ swarmcorp init
03$ swarmcorp start --project ~/code/your-app

Then run swarmcorp open (the dashboard picks a free port, 4317 by default) and start your coding agent in that project. Undo everything with swarmcorp uninstall.

Plate 07
Pricing

Free to run. Pay once for Pro.

No subscription. The core stays free and runs on your machine. Pro adds history and comparison, and pays for the work.
Free
$0
forever, local
  • ✓Live swarm map and timeline
  • ✓Context meters per agent
  • ✓Up to 12 terminals, plus live mirrors
  • ✓File-based dispatch queue
  • ✓Run receipts for the current run
Install free
ProShipping · v0.4
$29
one-time licence, one person, up to 3 machines
  • ◆Everything in Free
  • ◆Run history that survives restarts
  • ◆Receipt archive, with run-to-run compare
  • ◆Switch projects without restarting the hub
  • ◆Daylight theme
  • ◆A receipt saved automatically when the hub stops
Buy Pro
The Daylight theme: the same dashboard on a light background, with the timeline and all sub-tasks complete
Pro · Daylight theme
Receipt archive comparing two runs: tokens, agent runs, failed runs, average duration, tasks done and tokens per agent, with deltas
Pro · Receipt archive, run-to-run compare

Pay with Dodo Payments. Your licence key arrives by email and on the receipt page: paste it into Settings. It's activated once with Dodo, then re-checked about once a day, and keeps working offline for 30 days. No account. How Pro works.

Plate 08
From the log

Field notes on running agent fleets.

  1. A file queue, not a giant contextEngineering · 5 minWhy SwarmCorp agents talk through small JSON files instead of sharing one prompt: smaller contexts, atomic claims, a STATUS line on every turn, and feedback that goes straight back to the owner.
  2. Six agents, one laptopGuide · 5 minHow to run a CEO, a researcher and two implementers on a 16 GB laptop without the fan taking off: in-process subagents, background runs, idle sessions and MCP servers.
  3. Agent counts are cheap. Run logs aren't.Essay · 5 minA post says 400 agents shipped a product overnight. Here is how to check a claim like that, and what SwarmCorp writes down so your own numbers hold up.
  4. 12 workflow tips for Claude Code fleets
Plate 09
Questions

Before you install.

Does SwarmCorp send my code or prompts anywhere?+

No. The hub binds to 127.0.0.1 and has no account, no analytics and no cloud. It reads the hook events of your agents and the transcript files already on your disk.

Does it cost tokens?+

No. SwarmCorp never calls a model. It watches the sessions you run; the token numbers it shows are the ones your agents already recorded.

Which agents does it support?+

Claude Code, OpenAI Codex CLI, Gemini CLI, Google Antigravity (agy) and opencode. `swarmcorp init` connects every one it finds on your machine. Claude Code gets the most: launching role terminals and the dispatch queue are Claude-only for now.

Does it work with my subscription?+

Yes. SwarmCorp never touches authentication. It observes the CLIs you already use, logged in however you like.

What happens when the dashboard is off?+

Nothing you would notice. The hook relay gives up after about 0.8 seconds and always exits 0, so Claude Code never waits on it.

How does the Pro licence work?+

You buy once through Dodo Payments and get a licence key by email and on the receipt page. Paste it into Settings and Pro unlocks immediately. Lost the email? The billing portal on the Your licence page lists your key again. The key is activated once with Dodo (it sends the key and your machine name, nothing else), re-checked about once a day, and keeps working offline for 30 days.

Which platforms?+

macOS and Linux with Node 20 or newer. Windows is untested.

Is this an Anthropic product?+

No. SwarmCorp is independent and is not affiliated with or endorsed by Anthropic.