Docker images that turn Claude Code, OpenCode and Codex into Multica agents.
One container, four environment variables, and you have an agent runtime that picks up tasks, works in a real repository, and keeps its session history across restarts.
cp .env.example .env # MULTICA_TOKEN, AI_BASE_URL, AI_API_KEY, AI_MODEL
docker compose up -dNo CLI installation, no Node/PHP/Flutter toolchain on your laptop, no per-agent configuration. Every agent CLI is wired to the same endpoint from the same three values.
Running a Multica agent normally means installing an agent CLI, keeping it current, wiring it to a model endpoint, and repeating that on every machine. It gets worse when tasks need different toolchains: a Laravel task needs PHP and three databases, a Flutter task needs the Android SDK, a QA task needs Playwright and k6.
These images put the whole stack in a container that:
- ships Claude Code + OpenCode + Codex, all pointed at your AI router;
- registers with Multica automatically and runs the daemon in the foreground;
- runs as a non-root user (uid 1000) on Debian 13;
- keeps sessions on a volume, so a restart does not lose context;
- is built and tested in CI for
linux/amd64andlinux/arm64(Flutter isamd64only); - pins every tool in one file — no floating versions, no auto-update surprises.
One package, one tag per stack:
ghcr.io/proreka/multica-runtime:<variant>
| Tag | Toolchain | Use for |
|---|---|---|
base |
Node 26 LTS, Python 3, uv/uvx, git/gh, Chromium, build tools | general tasks |
node |
pnpm, yarn, TypeScript, Playwright, vitest, eslint | web frontends |
python |
Poetry, ruff, mypy, pytest | Python services |
php |
PHP 8.4, Composer, Laravel, MariaDB + PostgreSQL + Redis | Laravel/Symfony |
go |
Go, golangci-lint | Go services |
flutter |
Flutter, Android SDK, JDK 21, Gradle | mobile (amd64) |
devops |
Terraform, AWS CLI, kubectl, Helm, Docker CLI, Wrangler | infrastructure |
qa |
Playwright + MCP, axe-core, newman, k6, pytest | testing |
design |
ImageMagick, Playwright, Penpot MCP + plugin | UI/design |
Every variant includes Claude Code, OpenCode, the Multica CLI, git, gh, jq, and Chromium. See Variants.
Only four values are required:
MULTICA_TOKEN=... # https://multica.ai/settings?tab=tokens
AI_BASE_URL=https://... # Anthropic-compatible endpoint
AI_API_KEY=...
AI_MODEL=...Native variables still work and win when both are set:
| Public | Native |
|---|---|
AI_BASE_URL |
ANTHROPIC_BASE_URL |
AI_API_KEY |
ANTHROPIC_AUTH_TOKEN / ANTHROPIC_API_KEY |
AI_MODEL |
ANTHROPIC_MODEL |
The entrypoint picks Authorization: Bearer for routers and x-api-key for
Anthropic's own API, then generates the OpenCode config from the same values. See
Configuration.
tini
└─ variant entrypoint php: databases · design: Xvfb + MCP
└─ multica-entrypoint
1. resolve AI_* → ANTHROPIC_*
2. configure Claude Code + OpenCode + Codex
3. log in to Multica
4. exec multica daemon start --foreground
The daemon detects every CLI and registers one runtime per CLI. Sessions live on
/workspaces (.claude, .opencode/data, .codex). See Architecture.
| Target | Guide |
|---|---|
| Docker Compose | docs/deployment/compose.md |
| Docker Swarm | docs/deployment/swarm.md |
| Amazon ECS (Fargate) | docs/deployment/ecs.md |
| Kubernetes | docs/deployment/kubernetes.md |
| Dokploy | docs/deployment/dokploy.md |
The optional autoscaler scales each role between 0 and 1 based on pending agent tasks, with cooldown, grace, and running-task guards.
- Quickstart
- Configuration
- Variants
- Architecture
- Agent CLIs
- Persistence
- Autoscaling
- Troubleshooting
- Security
- Upgrading
- FAQ
The same content is published at https://proreka.github.io/multica-runtime/.
- Docker 24+ (Compose v2) or any orchestrator that can run a container
- A Multica token
- An Anthropic-compatible endpoint. OpenCode can also talk to an
OpenAI-compatible router; Claude Code still needs an Anthropic-compatible
surface.
AI_PROTOCOL=autopicks the right one for OpenCode. Codex needs an OpenAI Responses API (/v1/responses); setCODEX_BASE_URLifAI_BASE_URLdoes not serve it.
No GPU, no privileged mode, no Docker socket — except for the optional Swarm
autoscaler, which needs it to call docker service scale.
The agent runs arbitrary code by design. Treat the container as a sandbox for convenience, not as a security boundary against a malicious task. Scope the Multica and GitHub tokens accordingly.
- no secrets in image layers;
- non-root everywhere;
- no inbound ports;
- pinned, checksum-verified tool downloads;
- provenance and SBOM attestations on published images.
Read Security and report vulnerabilities privately per SECURITY.md.
Claude Code, OpenCode and Codex CLI are supported now. Qwen Code, Grok Build, DeepSeek Harness, Pi, Oh-My-Pi, Hermes Agent, OpenClaw and more are on the roadmap.
Issues and PRs are welcome. Start with CONTRIBUTING.md and
Development. CI builds only affected images; successful image updates on main automatically
create a patch release by tagging existing manifests, without rebuilding.
A local Docker daemon is optional.
