Skip to content
Atrium Atrium

Agents in Atrium

Atrium doesn’t wrap or proxy your agents — it runs them exactly as you would in a terminal, then layers visibility on top. What an agent gets beyond a plain terminal is declared by its provider.

The agent catalog

Section titled “The agent catalog”

New projects offer an agent group seeded from a user-editable catalog: Claude, Codex, OpenCode, Amp, Aider, and Gemini ship as presets (Settings → Tools). The catalog just names commands — listing an agent asserts nothing about whether it’s installed. Any command you add becomes a one-click agent template.

Providers

Section titled “Providers”

A pane’s provider is derived from its startup command’s first token (claude → the Claude provider). Providers declare capabilities; agents without any registered capabilities behave as plain terminals — which is always safe.

Claude Code is the fully-featured reference provider:

  • Session resume. Atrium mints a stable session id on first spawn (claude --session-id <uuid>) and resumes it on every wake, restart, and app relaunch (claude --resume <id>) — after checking the session still exists. You get the same conversation back, not a fresh agent.
  • Context-window %. A live chip shows how full the context window is — read from Claude Code’s own statusline, never estimated from a guessed window size. It needs Atrium’s one-line tee added to your Claude statusline; without it the chip simply stays hidden.
  • Managed scrollback. Claude redraws its own UI on resume, so Atrium skips scrollback replay for Claude panes rather than double-painting history.
  • Spawn environment. Claude panes are told not to auto-connect to an IDE, so diffs render in the pane you’re actually watching.

Other cataloged agents start generic, but any launcher whose CLI has a resume flag can be taught session resume — give its catalog entry create/resume commands (Settings → Tools → the agent’s commands) and its panes resume like Claude’s. The provider registry is how deeper integrations land without ever becoming control — a provider is a lens on an agent, never a driver.

A clean spawn environment

Section titled “A clean spawn environment”

Every pane — agent or not — is spawned “de-IDE’d”: VS Code integration variables are stripped, TERM_PROGRAM=atrium is set, a terminal editor is ensured for $EDITOR, and locale/color settings are normalized (especially in WSL). Agents behave as they would in a clean terminal, not as if they were inside someone else’s IDE.

Managed permissions (Claude)

Section titled “Managed permissions (Claude)”

Per project, Atrium can maintain a managed allow-list for Claude Code: rules you set in Project Settings are rendered into that project’s .claude/settings.local.json. The sync is deliberately non-destructive — your existing deny/ask rules, model choice, and env settings are preserved; Atrium only manages its own allow entries (deduped and sorted). Atrium never touches Claude’s notification settings; the waiting-pane breathe is derived from the pane’s own output stream (working → idle after a prompt you submitted), config-free and agent-agnostic.

TODOs and notes

Section titled “TODOs and notes”

Agents connected to the Atrium MCP server can still write TODOs and scratchpad notes into the project’s .atrium/ folder, but Atrium no longer displays them in-app (the Dock was retired in 0.15.0) — treat them as on-disk artifacts you read yourself.

Next

Section titled “Next”