Blog ·

Antumbra 0.2.0: available now

Relics, a gallery for the pages your agent writes, and behaviors that train into a standing expert and are checked again after they ship.

Antumbra 0.2.0 is available. It adds relics, a home for the pages your agent writes, and grows behaviors from rules you record into a loop that trains them into a standing expert and keeps checking that the expert still follows them.

Install or upgrade

Run the installer again to get the 0.2.0 antumbra command. No Rust needed.

# macOS and Linux
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/albedosehen/antumbra/releases/latest/download/antumbra-cli-installer.sh | sh

# Windows
irm https://github.com/albedosehen/antumbra/releases/latest/download/antumbra-cli-installer.ps1 | iex

To run the 0.2.0 server, pull your clone of the repository and run antumbra setup local again. Setup rebuilds the server from the clone and keeps everything already in place. MSI packages, archives and checksums are on the release page.

What’s new

Relics

  • An agent keeps an HTML page it wrote as a relic with one tool call, publish_relic. The page is stored in copal, and its text is indexed so recall finds it by what it says.
  • The dashboard has a gallery at /relics with a live preview of each page, filters for yours and shared with you, search, and a list view. Each relic opens at /relic/<slug>.
  • Publishing again under the same slug adds a version at the same address.
  • A relic is private by default, seen by you and whoever its compartment is shared with. Make one public and anyone who can reach your server opens it without signing in.
  • No address you type expires. A private relic is shown through a short-lived link made for each view.
  • A relic never runs on the dashboard’s own origin. It is shown in a sandboxed frame from copal’s render origin, with no network access of its own.

Behaviors you can teach (experimental)

  • behavior_candidates finds the rules you already stated in your memories, so you do not have to type them twice.
  • The dashboard has a Behaviors section where you accept or retire what was proposed.
  • Training holds one example in three out, and a behavior counts as learned when the expert passes enough of them. A behavior the expert answers worse than the base model is judged by how likely the drop is, not by a fixed margin.
  • After an expert ships, it is checked again on the examples its behaviors gain. An optional setting lets the loop tune its own confidence from those checks.
  • A behavior that was not learned shows what the expert answered instead, so you can see why.
  • The same rules train the same way on any machine, and answer can serve through your standing experts alone.

Recall and the server

  • A session that can read only a little of a workspace is searched exactly, in milliseconds instead of seconds.
  • A user id now names one user in one workspace, so a private expert can never be listed or routed to from another workspace.
  • The server answers at once after a restart: the first calls no longer wait on the chunk index or a cold embedder.
  • A page or document larger than 64 KB can be sent over the REST endpoint.

Docker and security

  • The Compose copal profile runs copal’s render listener, so a Compose install can keep relics.
  • The database health checks work on SurrealDB 3.2.4, and the deploy script reports a service that failed to start instead of saying it is fine.
  • Dependencies are past the October 2026 security advisories.

Known limits

  • Relics need copal with its render listener. The Compose copal profile runs it on your machine only by default, and a setting publishes it to your network.
  • Opening a private relic on another device means signing in there with your token. Sign-in by email link is not wired up yet.
  • Training experts still needs an NVIDIA GPU and a source build.

Get started, or read the full release notes.

← Every postFeed