Skip to content
kalyppsoPublic

About

A desktop workspace for running and supervising AI agents: sandboxes, secrets, budgets, policies and approvals.

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

KALYPSO logo: Kalypso drawn in fine wireframe lines, with coral in her hair

KALYPSO

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.

CI
Website · Download · Roadmap · Architecture

KALYPSO Today: quick actions for agents, tasks, sandboxes, secrets, policies and connections, above a year of commits by agents and people

Tour

A 40-second walk through KALYPSO, starting with kalypso handing two jobs to two agents.

KALYPSO tour: kalypso hands a flaky test and a README fix to two agents, then a walk through the app

Watch the full-quality video

What it does

Every call an agent makes goes through four layers before it reaches a model or a tool.

Layer What it does
Sandbox Each agent runs in a hardened container: read-only root, no Linux capabilities, its own network. Reaching a new domain can be made to need your approval.
Rules Policies in plain words, such as "never force-push to main", checked before every tool call. Each one blocks, asks you, or warns. Risky calls wait for your answer.
Vault Secrets are encrypted per namespace. Agents get a stand-in key, and the real one is added outside the sandbox. When an agent asks for a secret and you approve, only that one command gets it, and it is redacted from the output.
Gateway Every model call is metered per agent and per model, and stops at the budget you set. Works with Anthropic, OpenAI-compatible APIs and Ollama.

And around them:

  • Crew. Agents built on the Claude Agent SDK, supervised and restarted if they crash.
  • Namespaces. Separate networks, keys, budgets and policies for each workspace (personal, work, a client). Nothing crosses between them.
  • Links. Link agents to each other, to your tools (Obsidian, Notion, GitHub, any MCP server) or to other machines.
  • Observe. A tracker board, a Git tracker that tells agent and human commits apart, and a monitoring view.
  • Tools. Split terminals into sandboxes, a browser and a phone preview.
  • Code only when you ask. Configs, specs and diffs stay folded until you open them.

Screenshots

An agent's page: its model, sandbox, policies, skills, MCP servers, secrets and linked integrations Rules: block, ask me and warn rules in plain words, with suggested ones to add
Crew. What each one runs, where, and what it may touch. Rules. Block it, ask you first, or just flag it.
Vault: encrypted values with their scope and the agents that use them Spend: spend this month, daily spend per agent and budget rules
Vault. Encrypted, scoped, never shown to an agent. Spend. Spend per agent and model, with caps that pause or downgrade.

Board. What each agent is working on, with time, cost and the pull request it opened. Drag a task to change its status; click it to open the details beside the board.

Board with queued, running, in review, done and failed tasks

Split terminals. Shells into sandboxes, split right or down as many times as you like.

Four terminals in a two-by-two split, each connected to a sandbox

Browser and phone preview. A browser, and a phone-sized preview with common device sizes.

The browser pane showing a page The phone preview showing the same page on an iPhone-sized screen

Themes. Black, White and Blossom, or your own background, text and accent colours and font.

Today in the Black theme Today in the White theme Today in the Blossom theme
Black White Blossom

The screenshots show the browser preview on sample data. The desktop app starts empty and shows only your own agents, sandboxes and secrets.

Status

Pre-release (v0.1.x), working toward a public beta. In the desktop app today: namespaces, the encrypted vault and secret gateway, sandboxes on Docker or Podman with real shells, the agent supervisor with the Claude Agent SDK runtime, the policy engine with approvals, and the model gateway with budgets.

Next up for the beta: first-run setup, signed installers with auto-update, and agents linked to GitHub repositories. Progress is tracked in the Roadmap.

Download

Installers for Windows (.exe, .msi), macOS (.dmg) and Linux (.deb, .rpm, .AppImage) are on the Releases page and at kalyppso.github.io/website. Builds aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper will ask you to confirm the first launch.

Run it

You need Node 20+, Rust stable and the Tauri prerequisites for your OS. For sandboxes you also need Docker, or Podman with its API socket (systemctl --user enable --now podman.socket).

npm install
npm run tauri dev     # desktop app, backed by the Rust core
npm run dev           # UI only in a browser, on sample data

On WSL, if the window is blank, start it with WEBKIT_DISABLE_DMABUF_RENDERER=1.

Develop

cd src-tauri
cargo test                          # unit and IPC tests
KALYPSO_TEST_RUNTIME=1 cargo test   # also run real container tests (needs Docker or Podman)
cargo clippy --all-targets -- -D warnings
npm run check:contrast              # WCAG contrast of the built-in themes

TypeScript types for the IPC layer are generated from the Rust structs into src/bindings/ when the tests run. CI fails if they drift.

Contributing

Commits and pull request titles follow Conventional Commits: type(scope): subject.

feat(tracker): import tickets from Jira
fix(sandbox): detect Podman on Windows
docs(readme): add screenshots

Types: feat, fix, docs, style, refactor, perf, test, build, ci, chore, revert. Pull requests are squash-merged, so the PR title becomes the commit on main, and CI checks it.

Layout

src/            React UI (views, components, store, generated bindings)
src-tauri/src/  Rust core: store, ns, vault, sandbox, agent, policy, gateway, budget, events, terminal, ipc
docs/           Architecture and images

License

KALYPSO is open source under the MIT License.

About

A desktop workspace for running and supervising AI agents: sandboxes, secrets, budgets, policies and approvals.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages