Brick - Loop de Desenvolvimento Multi-Agent Autônomo
PhilipAndreasson/brick
Dev loop autônomo multi-agent local-first com escalação inteligente a Claude
Ver no GitHubAgent autônomo que orquestra um ciclo de desenvolvimento com múltiplos papéis (orchestrator, overseer, worker, smart_local, PO reviewer, wiki updater) rodando localmente em Ollama, escalando para Claude apenas quando necessário. Reduz custos de tokens ao maximizar execução local enquanto mantém qualidade através de gates de type-check e revisão.
Instalação
npx convez-ai PhilipAndreasson/brick1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de ciclos de desenvolvimento completos (design → código → review → commit)
- Redução de custos de API rodando tasks repetitivas em modelos locais antes de escalar
- CI/CD integrado via GitHub Actions com retry inteligente e auto-healing
- Desenvolvimento project-agnostic configurável via brick.config.yaml + PROMPT.md
Veredito

Brick é uma abordagem sofisticada e bem-pensada para reduzir custos de API mantendo qualidade através de orchestração local inteligente. Requer infraestrutura Ollama e compreensão de pipelines CI/CD, mas oferece retorno significativo em projetos que toleram automação de ponta a ponta com revisão humana opcional.
README
A tradução para o português deste README ainda está na fila.
Brick
Named after Brick Tamland. Loud, occasionally on fire, surprisingly effective.
Autonomous, multi-agent development loop for Claude Code + local LLMs.
Local-first. Brick runs six roles (orchestrator, overseer, worker,
smart_local, PO reviewer, wiki updater) on your local Ollama. Claude is the
top escalation tier: it only fires when local exhausts retries. Between the
workhorse local models and Claude sits smart_local — a stronger local model
that re-attempts hard tasks (repeated worker failures, overseer requeues, or PO
rejections) before any Claude tokens are spent.
The philosophy: 1000 monkeys with unlimited compute, not five monkeys with a tight deadline. Local hardware is free; Claude tokens are not. Throw attempts at the problem and let the type-check gate filter out the noise.
Brick runs in a tight loop on your workstation — or in CI via GitHub Actions —
and is project-agnostic: everything project-specific lives in your project's
brick.config.yaml + PROMPT.md.
Architecture (one diagram)
┌─────────────────────────────────────────────────────────────────────┐
│ brick.sh / brick.ps1 (loop, runs forever) │
│ │
│ At startup: spawn Overseer + N Workers (default = ~ CPU cores) │
│ │
│ Each iteration: │
│ 1. git pull --rebase --autostash │
│ 2. Idle check — fix_plan.md + tasks/pending + tasks/review │
│ 3. Local orchestrator (Ollama) ← reads PROMPT.md + fix_plan.md │
│ · picks one actionable item │
│ · emits OVERSEER or WORKER task JSON │
│ · moves the line under `## In Progress` │
│ 4. PO review (local Ollama, then type-check gate) │
│ · ACCEPT + tsc pass → done/, commit │
│ · REJECT or tsc fail → bump local_attempts, requeue │
│ · After local_attempts ≥ escalation threshold → Claude PO │
│ · After max_requeue_attempts → tasks/failed/ │
│ 5. Wiki update (local Ollama) from `git diff HEAD~1 HEAD` │
│ 6. Sleep, repeat │
└─────────────────────────────────────────────────────────────────────┘
┌────────────────────────────┐ ┌────────────────────────────┐
│ overseer.py (Ollama) │ │ runner.py × N (Ollama) │
│ │ │ │
│ Polls tasks/pending/ for │ │ Polls tasks/worker/ │
│ type:"overseer", expands │ │ pending/, builds prompts, │
│ into worker sub-tasks, │ │ hits Ollama, writes code, │
│ reviews completion, hands │ │ self-heals up to 3 │
│ off to PO review via │ │ attempts (or forever with │
│ tasks/review/. │ │ retry_forever:true), then │
│ │ │ escalates to Claude. │
└────────────────────────────┘ └────────────────────────────┘
Before Claude, every local escalation axis first promotes the work from the
workhorse model to smart_local (a stronger local model):
| Axis | Workhorse → smart_local after | → Claude after |
|---|---|---|
| Worker solving | worker_retries_before_smart (5) | 10 retries |
| Overseer plan/review | overseer_requeues_before_smart (2) | max_requeue_attempts (3) |
| PO review | local_attempts_before_smart (4) | local_attempts_before_claude (10) |
Claude only enters the picture once the smart-local tier is also exhausted:
- PO escalation — local PO has rejected the same review file
escalation.local_attempts_before_claudetimes (default 10). - Worker stuck — a single worker has retried 10 times on the same
task without success (and the task does not have
retry_forever: true).
Windows:
brick.ps1still runs the legacy Claude-orchestrator flow and does not yet implement the local-first ladder or the smart-local tiers. Usebrick.sh(Linux/Mac/WSL) for full local-first behavior. Seedocs/local-llm-stack.md.
Install
One-liner from your project root:
curl -fsSL https://raw.githubusercontent.com/PhilipAndreasson/brick/main/scripts/bootstrap.sh | bash
Pin a specific version:
BRICK_VERSION=v0.1.0 curl -fsSL https://raw.githubusercontent.com/PhilipAndreasson/brick/main/scripts/bootstrap.sh | bash
Or, from a local Brick checkout:
./install.sh /path/to/your/project
The installer drops .brick/bin/{brick,brick.sh,brick.ps1},
.brick/lib/{runner,overseer,config,init,doctor}.py, a VERSION file, and
the tasks/ + logs/ trees. It seeds PROMPT.md and fix_plan.md if
they are missing, but does not create brick.config.yaml — that's the
job of brick init (next step), so model choices stay explicit.
Quickstart
./.brick/bin/brick init # interactive setup → brick.config.yaml
./.brick/bin/brick doctor # verify claude / ollama / models / git / jq
$EDITOR PROMPT.md # describe your stack for the orchestrator
$EDITOR fix_plan.md # add the first item
./.brick/bin/brick run # start the loop (Ctrl+C to stop)
Other subcommands:
brick models # show current model assignments
brick model # arrow-key picker, swap any role
brick model set worker qwen2.5-coder:7b # direct set, no prompt
brick version # print installed Brick version
brick update # fetch latest release, preserve config
brick uninstall # remove .brick/, prompt before tasks/ + logs/
brick init --non-interactive \
--orchestrator gemma3:latest \
--overseer gemma3:latest \
--worker gemma3:latest \
--smart-local gpt-oss:20b \
--po-reviewer gemma3:latest \
--wiki gemma3:latest \
--claude-escalation claude-sonnet-4-6
brick model opens a curses TUI: arrow keys to pick the role, then to pick
the new model. Ollama options are pulled live from the daemon. Falls back to
a numbered prompt if there's no TTY.
Tip: add $PWD/.brick/bin to PATH (or symlink brick into
~/.local/bin) to call brick from anywhere inside the project.
Direct loop invocation (mostly for CI / scripts):
./.brick/bin/brick.sh --max-loops 5 # CI / time-bounded run
./.brick/bin/brick.sh --workers 0 # single Sonnet, no Gemma fleet
Windows: use pwsh ./.brick/bin/brick.ps1 — same flags, but it spawns
visible Terminal panes for overseer + workers instead of background log files.
Requirements
- Claude Code CLI (
claude) — for orchestrator + PO + wiki updater - Ollama running locally with at least one Gemma-class (or compatible)
model pulled — for the overseer + worker fleet. Brick talks to it at
http://localhost:11434(override inproviders.ollama.base_url). Pick the model atbrick inittime; there are no defaults. - Python 3.10+ (
pyyaml,requestsauto-installed on first run) - jq — used by
brick.shto manipulate review JSON - git
Run brick doctor to verify all of the above.
Config reference
See templates/brick.config.yaml.example
for every key with defaults.
Documentation
docs/architecture.md— full multi-agent designdocs/local-llm-stack.md— local-runtime reference: model→role mapping, engine choice, and slot/concurrency math for the RX 9070 XTdocs/brick-system.md— operational guidedocs/runner-tasks.md— task JSON schema, self-healing, Sonnet escalation
License
MIT