Agents that don’t stop
Sun is a self-hosted AI agent orchestrator. It turns the AI tools your team already pays for, like Claude Code, Codex and OpenCode, into workflows that keep going when a quota runs out and only count as done when the checks pass.
One binary · SQLite built in · your server or your laptop
Queued 2
-
October campaign drafts
-
Weekly summary
Running 1
-
Fix login timeout on mobile
Verifying 0
Needs you 0
Done 0
Fail an agent, route a task or pulse the core: the board answers.
// How it works
AI agent workflows in four moves
Hand it over. Watch it run. Trust the result, because the engine checked it.
Hand it off
Create a task from the board, Slack, Jira or the CLI. Start from a role template or build a workflow from blocks.
It runs
Step by step: agent, check, approval, notify. Each task gets its own workspace: a git worktree, or a task folder when there’s no repo.
It keeps going
Usage limit hit? Sun moves the task to another provider’s AI tool with a hand-off brief: plan, diff, last errors, to-dos. Or it pauses, or waits for the reset. You choose.
The engine checks
Tests, CI and content checks decide “done”, and they run again after a switch. Approvals really block. No check set? It says done (unverified).
// Try it
Watch a quota run out
Pick a task and press Run. This timeline is what you’d see when you open a task on the board, including the moment Claude Code hits its usage limit.
Claude Code starts a fix across two repos and hits its usage limit halfway. Codex finishes it from the hand-off brief, the tests run again, then you approve the PRs.
Press Run to start.
// For every role
AI workflows for every role
The founder or tech lead installs Sun once. Everyone else works from the web, Slack or a phone, with no git or terminal.
Role template
Bug fix → verified PR
When a task touches several repos, one task drives them all and comes back as linked PRs.
- Step: Plan
- Step: Implement
- Check: Tests and CI
- Step: Open PR
- Approval: Review
Role template
UI refactor with visual review
See the UI change, pin what’s wrong, then fix it yourself or ask the AI.
- Step: Implement
- Step: Design canvas
- Check: Checks
- Step: Open PR
- Approval: Review
Role template
Mobile feature, checked on a device
The agent builds the app and checks it on a simulator before it says done.
- Step: Implement
- Check: Verify on simulator
- Check: Checks
- Step: Linked PRs
- Approval: Review
Role template
Campaign: brief → drafts → approval
From brief to on-brand drafts to approval to scheduled posts, without chasing people.
- Step: Read brief
- Step: Channel drafts
- Check: Brand checks
- Approval: Approval
- Step: Schedule posts
Role template
Interview notes → PRD → tasks
Turn interview notes into a PRD, then into tasks you follow through delivery.
- Step: Read notes
- Step: Draft PRD
- Check: Section checks
- Approval: Approval
- Step: Create tasks
Role template
Weekly summary of work and cost
Progress, cost and blockers from your phone, plus a summary every Monday.
- Step: Gather runs
- Step: Summarize
- Step: Post to Slack
Role template
Client deployment from a template
One reusable workflow per client: their repos, data and plugins.
- Step: Load client setup
- Step: Configure
- Check: Checks
- Approval: Approval
- Step: Notify
// Building blocks, modded like a game
Snap on only what you need
A small engine that’s always on, plus mods: plugins connect the apps your team uses, extensions add what the engine can run or check. We ship some, and anyone can build the rest with the public SDK (Apache-2.0).
Engine · always on
Workflows, durable runs, watchdog, checks, failover, sign-in, run log
Plugins talk to someone else’s service
Extensions let the engine run or check something new
3of9 role templates ready
-
Bug fix → verified PR
Engineering
-
UI refactor with visual review
Needs Design canvas
-
Mobile feature, checked on a device
Needs iOS & Android simulators
-
Agent in its own VM
Needs VM or container
-
Campaign: brief → drafts → approval
Marketing
-
Interview notes → PRD → tasks
Needs Jira or Trello
-
Weekly summary of work and cost
CEO, CTO
-
Triage Jira and Slack into tasks
Needs Laya classifier
-
Instructions that learn from reviews
Needs Eval & auto-improve
// Runtimes and plans
Your AI, your keys, or ours
Mix all three in one failover chain. Bringing your own is always free, and there is no per-step billing.
Your subscription
Claude Code, Codex or OpenCode on the plan you already pay for.
Your API key
Anthropic, OpenAI, Google, OpenRouter, any OpenAI-compatible endpoint, or a local model on Ollama.
Sun credits
Start without either. Priced at cost plus a margin you see before each run.
Build your failover chain
Tap AI tools in the order Sun should try them when a quota runs out.
Failover chain: Claude Code, then Codex, then Sun credits.
You choose: switch, pause or wait. Checks re-run after every switch.
Self-host the full engine
- One binary on a VPS or your own Linux or Mac machine
- No seat or run limits
- SSO (OIDC) login and roles included
- Public SDK for your own plugins and extensions
- Source-available under FSL: read it, run it, change it
Or let us run it
- Sun Cloud: we host and upgrade it
- Sun credits for model usage
- Exclusive plugins and extensions built by Sun
- Support from the people who build it
Plans are still being set. No prices until they are.
// FAQ
Questions, answered
The short version of what Sun is, what it costs and where your data goes.
What is an AI orchestrator?
Software that hands tasks to AI agents, runs them step by step, and tracks and checks the results. Sun is a self-hosted AI agent orchestrator: it runs on your own machine.
Is Sun open source?
The engine is source-available under the Functional Source License (FSL): free to read, run and modify, and each release becomes Apache-2.0 after two years. The SDK and the runtime daemon are Apache-2.0.
Can I self-host Sun for free?
Yes. Self-hosting gets you the full engine, with no seat or run limits and SSO (OIDC) included.
Which AI tools does Sun work with?
Claude Code, Codex and OpenCode on your own subscription from v0.1. Your own API keys (Anthropic, OpenAI, Google, OpenRouter, any OpenAI-compatible endpoint, Ollama) from v0.2.
What happens when my Claude Code quota runs out?
Sun follows your policy: switch to the next AI in your chain with a hand-off brief, pause and notify you, or wait for the reset. Checks re-run after a switch.
How does Sun know a task is really done?
The engine runs checks: tests, CI and the pull request for code; content checks or an approval for other work. A task with no check shows as "done (unverified)".
Do I need git or a terminal?
Only whoever installs Sun does. Everyone else works from the web UI or Slack, with workflows built from blocks.
Does Sun send my code, data or keys to you?
A self-hosted install sends nothing out by default; telemetry is opt-in. Your API keys never reach Sun’s servers.
How much does Sun cost?
Self-hosting is free. Sun Cloud plans are still being set, so there are no prices yet.
What can I run it on?
Linux (x86_64 and arm64) and macOS (arm64).
Can I build my own plugins?
Yes, with the public SDK. Keep your plugins and extensions private or share them.
When can I use it?
Sun is in development. Design partners run it first: tell us one workflow you’d hand to Sun and we’ll be in touch.
// Design partners
Run Sun first
We’re picking design partners: small teams that already pay for Claude or ChatGPT and want them doing real work. You try Sun early and talk to the people building it.