Skip to content
brnrdbrnrd

Models & quota

brnrd separates the CLI process from the model it runs:

  • Shell: the CLI on PATH — claude, codex, grok, or vibe.
  • Core: the model and its cost, capability, and quota metadata.

Together they form the Runner for one wake. The resident remains the same when the Runner changes.

Inspect the profiles available on this machine:

Terminal window
brnrd runners list
brnrd runners list --all

Pin or let brnrd choose

Section titled “Pin or let brnrd choose”

Runner and daemon settings belong to the connected account, not to an individual repository. Inspect or change them from any connected checkout:

brnrd config show
brnrd config set runner.default codex
brnrd config set runner.default_class balanced
brnrd config set runner_policy fixed

These commands write <account home>/daemon.config. runner.default is an exact profile name; unset it with brnrd config unset runner.default and use runner_policy=cost-aware to let brnrd choose the cheapest adequate available local Runner. Tapping a profile in the dashboard rack sets the same account default and also parks that profile for the next wake, so the first wake and the wakes after it agree.

Older shell, core, runner, and default_class entries in .brr/config are read for one compatibility release. Daemon boot copies them to the account file, logs the migration, and leaves the old text in place but ignored.

Profile catalog

Section titled “Profile catalog”

Profiles are data in <account home>/runners.toml, with bundled defaults in src/brr/runners.toml. TOML keeps the operator-edited catalog human-readable while Python 3.11+ can parse it without another runtime dependency. An older account-home runners.md frontmatter catalog is read for one release and migrated at daemon boot.

Each [profiles.<name>] table may contain:

  • cmd: the headless command. The assembled prompt is piped on stdin unless {prompt} appears as its own argument.
  • binary: the executable to probe when the profile name is an alias.
  • hooks: the runner-specific Tier 2 hook adapter (claude, codex, grok, or vibe).
  • provider, owner, class, cost_rank, and quota_source: selection and quota metadata.
  • model: an optional pinned Core. The bundled Core registry also materializes model-specific profiles for catalogued Shells.

Core profile names are <shell>-<model slug> (codex-gpt-5.6-sol, claude-sonnet). The bundled [aliases] table keeps the former codex-mini, codex-terra, and codex-full pins resolving during the compatibility release; aliases do not appear as duplicate rack rows.

The minimum runner contract is a process that accepts the assembled prompt, operates in the supplied working directory, and exits with a status code. Printing a final reply on stdout adds response delivery; declaring hooks adds live tool-boundary injection. Profile commands and runner_cmd remain in the daemon-owned home because both decide which host command executes.

Grok Build

Section titled “Grok Build”

Install Grok with curl -fsSL https://x.ai/cli/install.sh | bash and sign in using Grok’s own setup (or set XAI_API_KEY). Select grok as a runner profile, or dispatch a strand with shell: grok. The adapter writes the wake to a temp file and passes --prompt-file, because Grok does not read a piped prompt. The resident protonucleus is appended with --rules, leaving Grok’s own tool instructions in place. A pinned core (grok-4.7, grok-4.6, grok-4.5, grok-4.7-build-fast) sets -m for that invocation only. The unpinned grok profile uses whatever model Grok itself would.

Headless runs use --permission-mode bypassPermissions and --trust. Grok loads the per-run .claude/settings.local.json (it treats that file as Claude-compatible hooks) and skips event names it does not know, so the post-tool seam is PostToolUse rather than Claude’s PostToolBatch. Pre-tool denies and stop blocks use Claude’s hook JSON.

The adapter prints Grok’s JSON envelope through. The runner reads it the way it reads Claude’s: the reply, the session id, the model that ran, the token totals, and the cost when Grok stamps one. A cost Grok omits stays omitted (pool and OAuth traffic often omit it; absence is not free, and per-model rows are not summed into one). A host-env seat resumes that session with --resume. Grok stores sessions under the working directory they ran in, so a worktree run does not arm a native resume. Quota and context-window headroom are not collected, and brnrd will not move a wake onto or off Grok because of a window it cannot see.

Mistral Vibe

Section titled “Mistral Vibe”

Install Vibe with uv tool install mistral-vibe and sign in using Vibe’s own setup. Select vibe as a runner profile, or dispatch a strand with shell: vibe. The adapter pipes the wake on stdin, installs an invocation-specific system prompt without replacing your settings, and returns the final assistant reply. It uses Vibe’s configured model and saved credentials (including VIBE_HOME).

Native boundary hooks are verified with Vibe 2.25.5: post-tool notices reach the model, file writes use the rooted-write guard, and post-agent denial requests a turn revision (Vibe caps retries at three). Each daemon invocation installs hooks in a temporary additional directory; BRR_VIBE_HOOKS=0 disables them. Session resume, model attestation and monthly quota/spend collection remain unavailable. Unknown capacity does not imply unlimited usage; Mistral subscriptions include a monthly usage allowance.

Escalate and downshift

Section titled “Escalate and downshift”

A resident can hand a hard continuation to a stronger local Core after it has read the repo, or pin an economy Core for bounded routine work. The handoff keeps the conversation and prepared worktree. Quality escalation does not silently opt into paid relay compute.

Quota is part of the live run posture and is shown to the resident before work starts. Automatic fallback is deliberately narrow: classified local quota/auth/provider failures may retry on another local Runner in the same or a cheaper class. Paid relay remains behind explicit consent.