Vancetope 0.5.0 is out 🎉

Vancetope 0.5.0 is tagged and pushed — 264 commits since 0.4.1.

0.4.0 connected Vancetope to the browser, to your keys, and to small models. 0.5.0 opens doors: agents that other agents can reach, an OAuth server that guards every door, a desktop app whose agents work on the machine in front of you — and, behind all of it, engines for the work that doesn’t fit into one chat turn.

Agents with a door

Trillian Next-Gen gives an agent a home: an identity, a hub, a model gate, schedules, and its own sessions. New in 0.5.0 are conversation heads — a Trillian can hold several conversations at once, each with its own head that reads, researches, and writes inside its hub — and the exposure door: an agents MCP at POST /brain/{tenant}/agents. Another agent (Claude Code, Cursor, or a second Vancetope) knocks, the gate decides, the conversation runs, and it closes when it goes idle.

Doors need locks. 0.5.0 ships a full OAuth 2.1 authorization server for the MCP doors: an issuer per tenant, consent for foreign clients, rotating refresh tokens with replay revocation, and dynamic client registration plus client-ID metadata documents. The CLI uses it too: /login --browser logs foot in through your browser, no password typed into a terminal.

And for the narrower case: a project MCP door at POST /brain/{tenant}/mcp/{project} exposes a fixed doc_* allow-list, scoped by a config document that says read-only or read-write and which paths.

Work that doesn’t fit into a chat turn

  • Wowbagger is the bulk engine. A Ford agent steers a mechanical worker pool over canonical JSONL input — thousands of records, one chunk publish at a time as the commit point, with wakeups for progress, failure cooldown, and heartbeat. Workers are pluggable (LightLlm, Zem, or a script), each behind a per-call model gate, and every run ends with a token cost report.
  • Zem brings decision models — System One for agents. A typed question goes in, a probability comes out, fast and cheap. zem_decide is a tool, the zem engine is spawnable, and command-safety is a bundled recipe that asks “is this shell command dangerous?” before it runs.
  • Wonko is a review loop: question sets, a review register, a check queue against reference documents. It routes free messages (an answer is not a new review), hands real work to a worker, and lets follow-ups re-enter the intake. Its questions arrive as the new questions document kind — a questionnaire with a chat wizard on the web and a line wizard in the CLI.
  • Nutrimat is a loop lab: one turn shell, many loop natures (janx, redbull, clubmate, absint, and a full coffee menu from ristretto to affogato), a loop judge, and narration of every round. It’s where we find out which loop shape works before it becomes a production engine.

A desktop app whose agents touch your files

The macOS desktop app (facelift) is no longer just the web UI in a window. Agents get client_file_* and client_exec_* on your machine — behind a sandbox policy, confined to a per-account workdir, with a master switch that says so loudly when it’s off. MCP tool packs pass straight through from ~/.vancetope/mcp.json, per project if you like, with credentials filled in in the app rather than in a config file. Live tabs per profile keep their state, ⌘1…⌘0 switch, and a two-finger swipe goes back.

The dmgs (Apple Silicon and Intel) hang on the release. They are ad-hoc signed, not notarized yet, so clear the quarantine flag once after installing: xattr -cr /Applications/Vancetope.app.

Settings, skills, and the rest of the surface

  • Settings panel: one guided place per scope — tenant, user, project — with setting forms, settings-doc kinds, and a raw editor, plus management areas for AI models, recipes, themes, engine prompts, scheduler, workflows, hooks, guards, wizards, mounts, and feed sources. Context help for every area. It replaces the old scopes page.
  • Skills now come and go: activation levels, an available-skills list, categories, and a turn lifecycle by default, so a task skill clears itself when the turn is done. Skills travel with delegated workers, and five coding skills ship in the box.
  • Scribble: handwriting sheets and a scribblebook — pen or finger, print, OCR and PDF export, and agents that can look at a sheet and draw on it.
  • Designer: design collections with a sandboxed live preview and a catalogue of design skills.
  • Hotblack: audio — text-to-speech, transcription, music, audio analysis, and local ffmpeg edits. Fenchurch got image_edit and an OpenRouter image provider.
  • Resource leases: a quota funnel from global down to user, defined as config documents and checked at use.
  • Chat: recipe-driven themes, an active message queue (keep typing while the engine works), ArrowUp history, and tool teasers in the progress stream.
  • Prompt caching is measured provider-agnostically, with a history breakpoint that keeps long conversations cheap.

Worth knowing before you upgrade

0.5.0 is a minor bump below 1.0, so it’s allowed to break. User hubs now get a home pod — on an existing database, switch the _user_* hub projects from HOMELESS to AUTO by hand. The OAuth client moved from /oauth to /oauth-client, because /oauth now belongs to the server. Docker images are on Docker Hub as 0.5.0 and latest; the full list, fixes included, is in the changelog. The CLI installs as brew install mhus/vancetope/vancetope; Java-free bundles hang on the release. See Upgrading for the walkthrough.

Try it

  • Get started — the Docker stack, a couple of commands, a few minutes.
  • Bring your own model — Vancetope ships no model; you connect your own.
  • Something broken? Discussions or an issue. Still early, still rough, still real.

Onward to 0.6. 🚀

🎉 more confetti