OmniAgent
CodeHawaii/omniagent
6 agentes de IA para código em um container Docker, zero credenciais pré-instaladas
Ver no GitHubImagem Docker multi-arquitetura com 6 agentes CLI populares pré-instalados (Claude Code, Codex, GitHub Copilot, Aider, Goose, Crush) em um único container. Construída sobre a stack omnidev com Python, Go, Node, C/C++ e sandbox hardened, sem incluir chaves de API — credenciais fornecidas em runtime.
Instalação
npx convez-ai CodeHawaii/omniagent1 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
- Executar múltiplos agentes de IA para codificação em CI/CD
- Ambiente isolado e reproduzível para desenvolvimento assistido por IA
- Testar e comparar diferentes agentes CLI (Claude, OpenAI, GitHub Copilot, etc)
Veredito

Solução madura e bem pensada para orquestração de agentes de código em containerização. A decisão de não incluir credenciais e suportar multi-arquitetura reflete design seguro; entretanto, a demanda por manutenção constante (evolução dos agentes upstream) pode requerer updates frequentes.
README
A tradução para o português deste README ainda está na fila.
omniagent
Popular CLI coding agents, pre-installed on top of the
omnidev toolbox image. One container
with Claude Code, OpenAI Codex, GitHub Copilot, Aider, Goose and Crush — plus the
full omnidev stack (Python/uv, Go, Node, C/C++, git, the hardened sandbox-run.sh).
docker pull ghcr.io/codehawaii/omniagent:latest
The image ships the CLIs, never any API keys. Provide credentials at runtime (env vars or a mounted config dir). See Authentication.
- Base:
ghcr.io/codehawaii/omnidev(Debian trixie, non-rootdev, multi-arch) - Arches:
linux/amd64,linux/arm64 - Run
agentsinside the container to list what's installed and which keys are detected.
Agents
| Agent | Command | Vendor | Install | Auth (headless) |
|---|---|---|---|---|
| Claude Code | claude | Anthropic | npm @anthropic-ai/claude-code | ANTHROPIC_API_KEY · claude -p |
| OpenAI Codex | codex | OpenAI | npm @openai/codex | OPENAI_API_KEY · codex exec |
| GitHub Copilot | copilot | GitHub | npm @github/copilot | GITHUB_TOKEN (Copilot-enabled) |
| Aider | aider | open source | uv (Python 3.12) aider-chat | any provider key (BYO) |
| Goose | goose | Block / Linux Foundation | official release binary | multi-provider |
| Crush | crush | Charm | official release binary | 15+ providers |
All six run headless and ship on
both amd64 and arm64. Current resolved versions are recorded in .versions.
Not included on purpose: Gemini CLI (Google is deprecating it 2026-06-18), OpenCode (archived → became Crush), Continue CLI (archived), Cursor CLI (proprietary, closed-source). Cline is omitted to keep the image lean (its CLI bundles ~0.8GB of vendored binaries).
Authentication
The image contains no credentials. Supply them at runtime, two ways:
1) Environment variables (best for CI/automation):
docker run --rm -it \
-e ANTHROPIC_API_KEY -e OPENAI_API_KEY \
-v "$PWD":/work -w /work \
ghcr.io/codehawaii/omniagent claude
2) Mounted config (best for local use — log in once on the host, reuse it):
docker run --rm -it \
-v ~/.claude:/home/dev/.claude:ro \
-v ~/.codex:/home/dev/.codex:ro \
-v "$PWD":/work -w /work \
ghcr.io/codehawaii/omniagent codex
Mount auth dirs read-only (:ro) so a misbehaving agent can't corrupt or
exfiltrate your tokens. agents prints which keys/config it sees.
Trust model — agents run code
These agents edit files and execute commands. Two modes:
Trusted repo (normal) Untrusted code (audit / review)
─────────────────────── ────────────────────────────────
your own project; agent wrap the agent in omnidev's
edits + runs with network sandbox-run.sh: read-only rootfs,
to reach the model API. caps dropped, resource limits.
`claude`, `codex exec`, ... `sandbox-run.sh --net <repo> 'aider ...'`
- For your own code, run the agent directly (it needs network for the LLM API).
- For untrusted third-party code, wrap it with the inherited
sandbox-run.shfor defense-in-depth. - Auto-approve flags (
--dangerously-skip-permissions, Codex--sandbox workspace-write, etc.) are convenient but should only be used on code you trust. The container is process isolation; the sandbox adds a second line of defense.
Quick start
# What's installed + which keys are set
docker run --rm ghcr.io/codehawaii/omniagent agents
# Claude Code, headless, on the current repo
docker run --rm -e ANTHROPIC_API_KEY -v "$PWD":/work -w /work \
ghcr.io/codehawaii/omniagent claude -p "summarize this codebase"
# Codex, non-interactive
docker run --rm -e OPENAI_API_KEY -v "$PWD":/work -w /work \
ghcr.io/codehawaii/omniagent codex exec "add a test for utils.parse()"
# Aider on a git repo
docker run --rm -e ANTHROPIC_API_KEY -v "$PWD":/work -w /work \
ghcr.io/codehawaii/omniagent aider --yes-always
Updating — upgrade all agents, cut a new image
The published image always builds the latest of each agent (Dockerfile ARGs
default to latest), so any rebuild upgrades them all.
./update-agents.sh # refresh ./.versions with current latest (BOM)
make update # same
gh workflow run publish.yml # manual rebuild → ghcr.io/.../omniagent:edge,:latest
git tag v$(date -u +%Y.%m.%d) && git push --tags # a dated, versioned release
make release # shortcut for the tag + push above
CI also rebuilds automatically every Monday (schedule cron in
publish.yml), so agents stay current without manual action. Pin a specific
agent at build time with e.g. --build-arg CLAUDE_CODE_VERSION=2.1.177.
Build & publish
make build # builds on omnidev:local -> omniagent:local
make agents # print bundled agents + versions
make run # build + interactive shell
publish.yml builds multi-arch and pushes to GHCR on push to main, v* tags,
manual dispatch, and the weekly schedule, with SLSA provenance + SBOM. PRs build
amd64 and smoke-test only (pr-build.yml).
The image builds
FROM ghcr.io/codehawaii/omnidev— that base package must be public (or readable by this repo) for CI to pull it. Making the GHCR package public is a one-time, irreversible step in the org package settings.
License
MIT — see LICENSE. Bundled agents retain their own licenses (Apache-2.0, MIT, FSL-1.1-MIT for Crush, etc.).