The guarantee

The #1 trust claim and the product's spine: the agent never sees your secrets. It is an enforced, verifiable invariant of the code — not a policy we ask you to trust.

The mechanism

A pasted key flows you → the daemon → libs/store, and never crosses the agent's MCP boundary. There is no MCP tool that returns a secret value — the agent's entire surface is key names + statuses + the wizard/action content it authors. When the daemon provisions, it makes the API calls with the stored creds and returns status, not values. There is literally no code path from a stored secret to the agent.

Why it holds even against a hostile agent

The agent (Claude, Codex, Cursor, Gemini) is capable but untrusted — assume it could be prompt-injected, compromised, or just buggy. The guarantee holds anyway because it is structural:

• No value-returning tool. The whole surface is plan / authorWizard / mapActions / renderWizard / renderActions / updateStatus / submitStep / executeStep / executeAction — every one returns names, statuses, or agent-authored content, never a value.
• Paste bypasses the agent. The value arrives at the daemon (POST /api/step, or submitStep), is validated and stored, and is never echoed back into any tool result, the SSE stream, or the browser response.
• Schema-validated UI. Every agent-authored Wizard / Step / Action is zod-validated; malformed is rejected. No freeform HTML/markdown as UI means no injection surface.
• Zero telemetry. The only network targets are the local daemon and the real provider APIs. No analytics, no phone-home, ever.

Enforced, not asserted

A leak-guard, check:no-leak, drives the full loop over MCP against the mock, captures every daemon → agent tool result and every SSE payload and the browser paste response, then asserts no sentinel value ever appears — while proving the loop really ran (the paste held its value, statuses flipped to synced). A leak fails the build, exactly like the boundary lint. The promise can't silently rot.

bun run check:no-leak

It's OSS — don't trust us, read it. Full threat model and how to audit in a few minutes: Security.