Kernl

0.4.4 · · Release on GitHub

Kernl 0.4.4

Upgrading from 0.4.3

  • The onion address no longer reaches the kernel. It used to forward to the kernel's own port, which put the dashboard, /mcp, /api/health and /api/metrics on a public address behind only the owner token. It now reaches a separate gateway that serves the presence card and nothing else, until you publish a service from the Friends page. Over Tor, friends reach only what you publish there.
  • Tor is started by the kernel, not by the container entrypoint. The switch on the Friends page turns it on and off without editing .env or restarting; KERNEL_TOR=1 only sets the default until the switch is used. The onion key in /app/data/tor/kernl is kept, so the address does not change.
  • Choose an edition. Setup gains a personal/business step. On an install from before editions (migration 4 marks it), choosing one only ever turns extensions on, never off.

Added

  • Personal or business edition. It decides which free extensions start active, which starter team setup offers, and the context every agent gets. Change it from Settings → General; the extension store filters by it and marks extensions made for only one.
  • Model Bench (/bench) scores every connected model on Kernl's own kinds of work: tool calls, agent tasks, JSON, numbers, language, long context and prompt injection, with deterministic checks, cost and history. The LIVE pill shows one dot per model of the chain and each one's score.
  • System overview (/control): how this Kernl is doing at a glance, with a way into every part of it.
  • "What would the AI answer?" on an office question is now answered by a read-only investigator that reads the issue, its merge request, the code and the git history before giving an opinion, and shows its progress.
  • Office panels. An office can declare panels (sources, views and actions, panels.v1) in the Office Kit manifest or an extension's office seed. They show as tabs in the office's 3D panel and open from the whiteboard in its room. Every read and action is validated, rate-limited and audited.
  • More AI providers: Cerebras, Mistral, Together, Fireworks, Kimi, Qwen, Z.ai, Xiaomi MiMo, and any OpenAI-compatible server. They come with brand icons and a two-column picker.
  • Tor in the native packages. Linux, macOS and Windows builds ship Tor, so a native install can be reachable from anywhere too.
  • Answer a report's choices from My Office; the pick goes to the agent's chat.
  • An agent can opt out of self-grading after each run, and extension seeds can opt one out.
  • scripts/kernl.sh starts, stops, tails and checks the local stack.
  • A compact JSON viewer for tool calls, results and agent messages.

Changed

  • Scheduled agents run in parallel, so one long run no longer holds back every other schedule.
  • A connected provider's model is chosen with the model picker, Claude Code included; the model is tested before it is saved.
  • The message stream shows each post's time and tokens in their own columns on the left: the clock, how long the post took, the tokens it consumed and the run total so far. claude_code agents now report their tokens per model turn instead of only at the end of the run.
  • The office panel is wider, its board is coloured by state, and Destroy office moved out of the footer into a danger zone.
  • 3D walkers take the shortest corridor path around desks, meeting rooms and the reception, and carry their message inside their nameplate.
  • The Claude sign-in redirect is caught on the host, so a Docker stack on Linux needs no paste.
  • The chief's unresolved questions reach the operator as soon as its triage run ends, and unattended runs reach the operator through the chief.

Fixed

  • OpenAI's reasoning models (GPT-5, o-series) refused every request: max_tokens is now sent as max_completion_tokens, unsupported sampling is dropped, and their token cap leaves room for the reasoning.
  • The model picker in the provider dialog opened underneath it.
  • claude_code agents' long replies were cut in the chat thread.
  • The camera missed the office it was focusing at some frame rates.
  • A redrawn provider logo stayed hidden behind the browser cache.
  • A sandboxed claude_code run could rotate Kernl's Claude session away, leaving every later run with "OAuth expired". Sandboxes now get the access token without the refresh token, the kernel refreshes it first when it would expire mid-run, and a dead session counts as an auth failure.