KALYPSO Product How it works FAQ GitHub Download

Run AI agents safely,
on your own machine.

Each agent works in its own sandbox. KALYPSO checks every tool call against your rules, keeps your real keys out of reach and caps what it spends.

Download Star on GitHub
kalypso Claude Sonnet 5.5Ready

What should we work on?

    Crew / builderRunning
    TaskFix the failing config test and open a pull request.
      Rules
      • No force-push to mainblock0 today
      • Ask before unknown hostsask me2 today
      • Never reveal secretsblock0 today
      • Stay inside /workblock0 today
      • Stripe only for billing testssecret requests1 today
      Playbook
      • Careful reviewerpersonareviewer
      • Tests firstinstructioncoder, builder
      • Short reportsinstructionevery agent
      • Ask before migrationsinstructionkalypso

      Write a prompt once and give it to any agent. Each gets its persona and instructions when it starts.

      Spend

      work, today$0.84 of $5.00

      work, this month$31.20 of $100.00

      client-acme, this month$19.40 of $20.00

      At the cap, calls stop at the gateway. They never reach the provider, so they cost nothing.

      Vault
      • ANTHROPIC_API_KEYmodel gateway onlynever in a sandbox
      • GITHUB_TOKENfile/run/kalypso/secrets
      • STRIPE_KEYon request1 approved command
      • DATABASE_URLfileapi sandbox

      Agents only ever hold a revocable kx_ key.

      Works with the tools you already use

      Ask once. kalypso runs the crew.

      Tell kalypso what you need. It plans the work, starts an agent for each part, runs them in parallel and reports back, with every step open to read.

      • Plans and delegates. Creates, starts and stops agents for you, inside your rules and budget.
      • Remembers. Keeps what matters across chats, and you can remove anything.
      • Any model. Claude, or any OpenAI-compatible model, local ones included.

      What you can hand off.

      The work you'd rather not babysit. kalypso splits it up, your crew does it, and you review the result.

      Fix the failing tests

      It reproduces the failure, writes the fix on its own branch and opens a pull request.

      Review pull requests

      A reviewer reads every change and flags what blocks a merge.

      Research and write it up

      Sources read, findings summarised, links included.

      Watch the deploy

      An agent keeps an eye on the release and tells you the moment something breaks.

      Paintings: Turner, Tischbein, Draper and Waterhouse, from the Odyssey. Public domain, via Wikimedia Commons.

      Every call goes through every layer.

      From your agent to the model, KALYPSO is in the path, and it asks you before anything risky.

      01SANDBOXIts own box and network.02POLICIESEvery tool call checked.03VAULTReal keys added here, not inside.04GATEWAYMetered and capped per agent.AGENTAnthropicOpenAI-compatibleOllamablockedasks you

      Point at a layer to see what it does.

      Every agent, at work.

      What each one runs, where it runs, and what it may touch.

      Rules, in plain words.

      Block it, ask you first, or just flag it. Checked before every tool call.

      Every agent in a box.

      Read-only containers with no privileges, each on its own network.

      Secrets stay in the vault.

      Agents get files or a stand-in key. Never the real value.

      Spend has a ceiling.

      Every token metered at the gateway. Calls stop at the cap.

      An agent called coder running in its sandbox, with its model, policies, skills and secretsAn agent called coder running in its sandbox, with its model, policies, skills and secretsSupervised: restarted if it crashes.
      The Policies page: block, ask me and warn rules, with today's stopped callsThe Policies page: block, ask me and warn rules, with today's stopped callsBlock, ask you, or warn.
      The Sandboxes page: two isolated containers with their limits, network and statusThe Sandboxes page: two isolated containers with their limits, network and statusStarted, paused or stopped from here.
      The Secrets page: encrypted secrets with their scope and which agents use themThe Secrets page: encrypted secrets with their scope and which agents use themValues never reach an agent.
      The Budget page: spend this month, daily spend per agent and budget rulesThe Budget page: spend this month, daily spend per agent and budget rulesMetered per agent, per model.

      Secrets stay put.

      Keys live encrypted on your machine. Agents get a revocable stand-in, and approved commands get the real thing without the agent ever seeing it.

      Every call is checked.

      Rules in plain words run before each tool call. Risky ones wait for you. If KALYPSO can't answer, the answer is no.

      Spend has a ceiling.

      Every token is metered at the gateway. At the cap, calls stop before they reach the provider.

      Free. Open source. Yours.

      No account, no telemetry. Needs Docker or Podman.

      Roadmap · How it works · Changelog · Code signing policy

      Questions

      Which models?

      Claude through the Anthropic API, and any OpenAI-compatible endpoint, including local ones like Ollama.

      Does my code leave my machine?

      Only the model calls your agents make, to the provider you choose.

      What do I need installed?

      Docker or Podman, and an API key for your model.

      macOS won't open it.

      Early builds aren't notarized yet: System Settings, Privacy & Security, Open Anyway. Once.

      Windows warns about the installer.

      Early builds aren't signed yet: More info, then Run anyway.