v3Coding agent — repo chores, PRs, secret rotations.
Open source · Apache-2.0
Orrery is the platform that keeps AI agents in motion — sessions that survive crashes, sandboxes for their hands, vaults for their keys, and a harness you can swap. We run it for you, or you run it yourself.
Anthropic-compatible API · SQLite, Postgres, or D1 · Managed or self-hosted
The console
A live sketch on autopilot — it mints keys, creates agents, and kills its own processes. Move your cursor in to take over.
Install the CLI, mint a key, and let agents operate against your control plane.
The command-line client lets your agent (or you) drive the platform from the terminal.
npx -y -p @orrery/cli orrery
or globally: npm i -g @orrery/cli
Your agent needs this to authenticate. Keep it somewhere it can read.
Point your agent at the CLI or API skill, then ask for what you want:
Use Orrery to create a research agent that monitors arXiv daily
Open to watch a live agent — then kill it.
sess_01hxq4kj
running
v3Coding agent — repo chores, PRs, secret rotations.
v7Linear teammate — reproduces, comments, tags fixes.
v2Scheduled — reads support inbox, ships a morning brief.
State that survives a crash. Secrets that never touch a sandbox. Somewhere for tools, skills, and memory to live. Orrery is that everything — the machine around the model.
How it works
Model, system prompt, tools, skills, MCP servers. Versioned, like code.
An append-only event log is born. Stream it live over SSE.
Subprocess, Firecracker, Daytona, E2B, or a Cloudflare Container.
Tools run, vaults sign requests at the network layer, memory mounts as files.
A process, a container, a region. It happens. That's the design assumption.
State rebuilds from the event log — mid-thought, same place. No luck involved.
Features
Every turn is an append-only event. Kill the process mid-thought — it resumes from the log, not from luck.
Subprocess for speed, Firecracker for isolation, Daytona, E2B, or Cloudflare Containers — one interface.
Credentials are injected at the network layer. env | grep TOKEN returns nothing — a prompt-injected agent has nothing to leak.
Per-tenant LLM credentials, encrypted at rest. Anthropic, OpenAI, or any compatible endpoint — rotate without a redeploy.
Instruction packs mounted into the sandbox, and memory as plain files the agent reads with its own hands.
Publish agents into Linear, GitHub, and Slack under their own identities — assigned, mentioned, replied to.
For your agents
$ npx -y -p @orrery/cli orrery login
✓ signed in — control plane: console.orrery.sh
$ orrery agents create scout --model claude-sonnet-4-6 --toolset default
{"id":"agent_01hx…","name":"scout","version":1}
# kill the process. redeploy. the session resumes mid-thought.
Deploy
We run the control plane, sandboxes, and upgrades. You bring the agents.
Book a demo →Workers, Durable Objects, and Containers — in your own Cloudflare account.
wrangler deploy
Your VPS or Kubernetes, SQLite or Postgres. One container and a URL.
railway up
Built in the open, Apache-2.0 end to end. Read the code, swap the harness — or let us run the whole machine for you.