Brain jobs on your subscription

Living Brain can use a Claude Pro/Max or ChatGPT Plus/Pro subscription as the model for the brain's jobs (compiling your coding-agent sessions, answering Ask) — but only the way those subscriptions are allowed to be used: inside the vendor's own client, on your own machine.

The rule this is built around

Anthropic and OpenAI permit consumer subscriptions only inside their own clients (Claude Code, Codex). So a subscription's credentials — the Claude.ai OAuth token, the "Sign in with ChatGPT" token — are never sent to, stored on, or used from the Living Brain server. Instead:

1. You choose "run my jobs on my device" (Settings → Models, the Claude subscription or ChatGPT subscription card).
2. When one of your jobs needs a model, the server queues it instead of calling an API. The job holds the same redacted prompt an API model would have been sent.
3. livingbrain jobs run (or the Mac app, which runs it for you) claims the job, runs your claude or codex with the prompt, and posts the text back. The server redacts and validates the answer again on arrival.

Only the job's owner can claim it; a claim is a lease, and a job whose runner went away becomes claimable again when the lease lapses. Tokens the CLI reports are recorded as usage at no cost to the workspace.

Which jobs, and how they wait

• Compiling your coding-agent sessions (docs/coding-agents.md): each model call becomes a job and the compile shows waiting-runner. When your runner posts the answer, it must parse as the compile's closed edit list or the job fails (model-runner-bad-output) and the compile retries. A valid answer continues the compile in the same request, from the chunk where it stopped.
• Ask (POST /v1/ask): retrieval runs as always, over your own scopes. The answer is then written by your runner instead of the workspace's model, and the route returns 202 {"status": "pending", "job", "poll", "retry_after", "citations", …}. The client polls GET /v1/model-jobs/{job} until it leaves queued/claimed, then sends the same ask again. That second ask reads the runner's answer, redacted and citation-checked exactly like a model's (mode: "model", model.provider: "local-runner:claude"). livingbrain ask, the web app and the Mac app all wait up to 5 minutes and say what they are waiting for. Because the answer costs the workspace nothing, the monthly model budget does not gate it.

What runs

The prompt (the job's user turn) goes on stdin. Each job runs in a fresh, empty, private temp directory that is deleted afterwards, with the Living Brain token removed from the child's environment. Verified against Claude Code 2.1.300 and codex-cli 0.148.0:

# Claude Code — no tools, no MCP servers, no settings, no slash commands, one turn
claude -p --output-format json --tools "" --permission-mode dontAsk \
  --strict-mcp-config --setting-sources "" --disable-slash-commands \
  --no-session-persistence --max-turns 1 --system-prompt <instructions> \
  [--model <sonnet|opus|haiku|fable|full id>]

# Codex — the read-only sandbox, ephemeral, user config and rules ignored;
# stdin is the instructions, a blank line, then the prompt
codex exec --sandbox read-only --skip-git-repo-check --ephemeral \
  --ignore-user-config --ignore-rules --json --color never \
  -C <tmpdir> -o <tmpdir>/last-message.txt [-m <model>] -

--setting-sources user replaces "" only for a job set up as "Claude Code pointed at Z.AI's GLM Coding Plan", which lives in your own Claude Code settings. No permission-bypass flag (--dangerously-skip-permissions, bypassPermissions, --dangerously-bypass-approvals-and-sandbox, danger-full-access) is ever passed; the CLI refuses to run a job whose command line would hold one.

Nothing of a prompt, an answer or the CLI's own output is logged. A failure is reported (locally and to the server) as one of timeout, cli-missing, not-logged-in, cli-error, bad-output.

Commands

livingbrain jobs agents            # which CLIs are installed and signed in
livingbrain jobs run               # run your jobs until Ctrl-C (one at a time)
livingbrain jobs run --once        # at most one job, then exit
livingbrain jobs run --agent codex # use Codex for every job
livingbrain --json jobs run        # one JSON event per line (the Mac app reads these)

Options: --poll <s> (default 5) between polls when idle, --timeout <s> (default 300) per job, --claude <path> / --codex <path> when the binary is not on PATH. A signed-out Living Brain token exits with status 2.

--json events: started, claimed (job, kind, agent, model), done (job, kind, seconds, prompt_tokens, completion_tokens), failed (job, reason), error (message; the runner keeps going with backoff), stopped.

Headless (a server, colony1, omarchy, any Linux box)

livingbrain login                 # device flow: approve in any browser
claude auth login                 # or: codex login — on this machine, once
livingbrain jobs run --agent claude

As a systemd user service:

# ~/.config/systemd/user/livingbrain-jobs.service
[Unit]
Description=Living Brain jobs on my Claude subscription
After=network-online.target

[Service]
ExecStart=%h/.local/bin/livingbrain jobs run --agent claude
Restart=on-failure
RestartSec=30

[Install]
WantedBy=default.target

systemctl --user daemon-reload
systemctl --user enable --now livingbrain-jobs
journalctl --user -u livingbrain-jobs -f   # job ids, timings and token counts only

On a headless Linux box the CLI's token lives in the Secret Service keychain; where there is none, set LIVINGBRAIN_TOKEN in the unit's environment from a personal access token (Settings → Tokens).