Components & architecture¶
antcrew is made of four independent pieces. Each can run without the others, but together they give you end-to-end observability and control over your AI pipelines.
antcrew — the agent framework¶
antcrew is the Python package your team builds and ships agent logic with.
What it does:
- Provides ready-made agent roles — PM, developer, reviewer, QA — built on LangGraph
- Ships a CLI (
antcrew run …) to execute pipelines locally or in CI - Integrates with Slack, Telegram, and other notification channels
- Depends on
antcrew-enginefor all LLM calls, tracing, and HITL logic
What it is not: a server or a UI. It's a library you install in your project.
antcrew-engine — the core SDK¶
antcrew-engine is the low-level Python library that the framework builds on. You can also use it directly without the full agent framework.
What it does:
- Turns Python functions into LLM prompts via
@contract— the docstring becomes the system prompt, the type annotations enforce the output schema - Runs every LLM call through a configurable provider (
openai:gpt-4o,anthropic:claude-opus-5,simulated:echofor tests…) - Writes every token, tool call, and intermediate result to a TraceLog — a structured, append-only record of the full run
- Exposes
hitl_checkpoint()— a blocking call that pauses agent execution and waits for a human to approve before continuing
What it is not: a server, a database, or a UI. It's a library.
antcrew-platform — the cloud control plane¶
antcrew-platform is the managed web application running at antcrew.org.
What it does:
- Receives runs from the engine via
POST /runs - Stores every event — status changes, TraceLog events, tickets created — and serves them over a real-time WebSocket
- Serves the dashboard so your team can watch runs live, review HITL queues, and inspect tickets
- Manages HITL review queues — reviewers see pending approvals and can approve, reject, or comment
- Extracts structured tickets from run output with workspace-scoped display IDs (
PROJ-00001) - Sends outbound webhooks to your own systems when runs complete or reviews are needed
What it is not: it never executes your agent code. It is an observer and gating layer, not a worker.
antcrew-proxy — the LLM gateway¶
antcrew-proxy is an OpenAI-compatible HTTP proxy.
What it does:
- Accepts model calls from the engine using the standard
POST /v1/chat/completionsinterface - Looks up the caller's workspace in the platform and injects the right provider API key (BYOK)
- Routes to the correct upstream provider based on the model prefix in the request (
openai:,anthropic:,groq:,gemini:…) - Your application code never handles LLM credentials directly
What it is not: required. You can use the engine without the proxy by setting provider keys directly on the Agent. The proxy is valuable when you have multiple workspaces with different LLM budgets or key rotation requirements.
How a run flows through the system¶
sequenceDiagram
autonumber
participant Code as Your agent code<br/>(antcrew / antcrew-engine)
participant Proxy as antcrew-proxy
participant LLM as LLM provider
participant Platform as antcrew-platform
participant Human as Human reviewer
Code->>Platform: POST /runs — start a new run
Platform-->>Code: run_id
Code->>Proxy: POST /v1/chat/completions (model call)
Proxy->>LLM: forward with injected API key
LLM-->>Proxy: completion
Proxy-->>Code: response
Code->>Platform: POST /runs/{id}/events — TraceLog entry
Note over Code,Platform: hitl_checkpoint() reached
Code->>Platform: POST /reviews — pending review
Platform->>Human: notify (webhook / email)
Human->>Platform: approve
Platform-->>Code: resume signal
Code->>Platform: POST /runs/{id}/events — TraceLog (run complete)
Component overview¶
| Component | Role | Where it runs |
|---|---|---|
antcrew |
Agent framework + CLI | Your codebase |
antcrew-engine |
LLM calls, contracts, TraceLog, HITL | Your codebase |
antcrew-platform |
Dashboard, storage, HITL reviews | antcrew.org (managed cloud) |
antcrew-proxy |
LLM routing, BYOK key injection | antcrew.org or your own infra |