Skip to content

MITHOS Cyberdeck

A sovereign, single-operator command console for working with AI agents — where the model, the agent, its permissions, every tool call, and every dollar are visible and under your control.

Status: prototype. The console, its state model, approvals, ledger, and persistence are real. Models, tool execution, and Lab environments are simulated — clearly labelled in the UI. No backend, no accounts, nothing leaves your browser.

Try it live · Mission Control (HELM): /#helm


Why

Chat UIs hide the part that matters: which model answered, what the agent is allowed to do, what command it wants to run, and what it cost. MITHOS puts those on screen as first-class objects and makes every one of them an operator decision:

  • Sessions → Runs → Tool calls → Events. An agent proposes an exact command (executable, arguments, working directory, reason, expected effect, risk, expiry). You approve once, reject with a reason, or stop. Everything is an append-only event you can replay later.
  • Manual or Auto. Auto approves low-risk tool calls on arrival and logs that it did; anything else still asks.
  • Budget on screen. A context-usage ring, per-session model choice, and a streaming speed control (Fast / Medium / Thorough) that actually changes behavior.
  • Explorer. Index your own machine (read-only script), search everything, and find byte-identical duplicates — because "where did that file go" is half of using these tools.
  • Lab. Paste a GitHub URL → inspect what it needs and its risk → isolate → run a bounded smoke test under budget → adopt or destroy. Trying is free; trusting is a decision. Read the design in docs/LAB_TRY_MECHANISM.md.
  • / commands. /try owner/name, /model remote, /speed fast, /auto, /attach, /stop
  • HELM / Mission Control. A second screen: one mission, three competing branches (primary / adversary / alternate), an agent rail, Freeze Writes and All Stop, an event ledger.

Run it

npm install
npm run dev          # http://localhost:5173   — HELM at /helm
npm test             # 37 tests: state transitions, approvals, persistence, slash commands, Lab
npm run build        # tsc -b && vite build

Index your own files for the Explorer (read-only, output is gitignored):

node scripts/index-artifacts.mjs
# optional: scripts/explorer-roots.json to choose which folders — see the script header

Design system

Dark, dense, hard-edged. Built on MITHRAEUM tokens (src/styles/mithraeum.css): near-black surfaces read by light not hue, bone as the single content accent, square corners, monospace for data. A white tier (src/styles/mithos-white.css) is reserved for the top CTA and structural separators; yellow means health, never decoration. Contribute within it — see CONTRIBUTING.md.

Roadmap (short version)

  1. Real Inspect — GitHub API + manifest/install-script parsing + OpenSSF Scorecard, read-only.
  2. Real isolation — Docker / VPS / sandbox backends with observed egress and filesystem reach.
  3. Cost-governance spine — estimate → budget check → gate, structurally before every call.
  4. Multi-box, secrets broker, connector parity.

Long version and locked architecture decisions: docs/.

Not affiliated

MITHOS is an independent project. It is not affiliated with, endorsed by, or associated with Anthropic, OpenAI, Nous Research, or GitHub. Some interface conventions are deliberately familiar; the product is not theirs.

License

Apache-2.0.

About

Sovereign single-operator command console for AI agents — visible models, permissions, tool calls, cost. Prototype.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages