Bots that have to prove "done" before you believe them

What it does

kwant packs a multi-agent workspace into a single self-hosted binary: a Rust backend (Tokio, Axum, SQLite) with an embedded React web app. You get bots with individual roles, models and “souls” that @mention each other to hand work off, plus channels, scheduled routines, an Inbox for what needs you, and separate spaces for work and home. It runs with Claude, Codex or Grok CLI logins, your own API keys, or an OpenAI/Anthropic-compatible server, and rotates accounts when one runs dry.

The interesting bit

The core idea is distrust as a feature. A task only turns Verified when kwant runs the check itself — a test, a file check — and keeps the receipt; a bot’s word or attached files count as evidence, never proof. In 0.9, a second bot on a different model re-runs checks in a clean copy of the files, and for high-risk work it mutates the code slightly to see whether the tests actually notice. Payments always require approval, in every mode, even behind a plain “Continue”.

Key highlights

• Mission board turns each hand-off into a card with a verdict, evidence and cost per verified result
• Routing is transparent: each message names the bot that answered and why, with a “Wrong bot?” correction path
• Web facts must link to a page containing the quote word for word, or kwant refuses them
• Approval system shows the exact command, with per-space/per-bot modes and a Flight School trust ladder you grant manually
• No telemetry; the interface loads nothing from external servers (the optional Laya router is the one exception, and --no-laya disables it)

Caveats

• Linux is marked beta and Windows beta-er still; macOS is the first-class citizen
• The feature list is enormous — dev worktrees, IMAP loose-ends, Parallel Universe deploy comparisons — which suggests a lot of surface area for a young project (508 stars) to keep polished
• Bot budgets are currently set through the API only

Verdict

Worth a look if you want agent autonomy without taking a bot’s “done” on faith, and you’d rather self-host than rent it from xAI or OpenAI. Skip it if you wanted a minimal single-agent chat wrapper — this is a full operating environment, and that’s the point.