antcrew

Multi-agent framework for Python. Typed outputs. Full trace. Works offline.

Three lines to your first agent team, no API key required:

from antcrew import QuickStart

result = QuickStart.dev().run("Build a FastAPI auth service")
print(result.state["prd"].title)           # typed PRD artifact
print(result.state["code_artifacts"])      # typed code files

Or from the CLI, fully local with Ollama:

pip install antcrew
antcrew run "Build a FastAPI auth service" --model ollama:llama3

Why antcrew

The core differentiator: every output is a typed artifact (Pydantic class, not a dict) and every decision is recorded to a local TraceLog you can replay. Both work offline, for free.

Quick start

Zero setup — simulated LLM, no credentials:

Runs immediately. Produces typed artifacts with deterministic fake content — good for testing and CI, not for real AI output.

pip install antcrew
antcrew run --model simulated "Build a REST API for user authentication"

Fully local — Ollama (real AI, no API key, no data leaves your machine):

Requires Ollama installed (~5 min) and ollama pull llama3.

antcrew run --model ollama:llama3 "Build a REST API for user authentication"

Cloud model — real AI, no local setup:

export ANTHROPIC_API_KEY=sk-ant-...
antcrew run --model claude "Build a REST API for user authentication"

Don't want to configure anything? Use antcrew-platform — the managed tier provides the LLM. You run agent teams from a web UI without installing Ollama or managing API keys.

From Python:

from antcrew import DevTeam
from antcrew.models import OllamaModel, AnthropicModel, SimulatedLLM

# Local — no API key
team = DevTeam(model=OllamaModel("llama3"))

# Cloud
team = DevTeam(model=AnthropicModel("claude-sonnet-4-6"))

result = team.run("Build a REST API for user authentication")
print(result.state["prd"].title)           # PRD object
print(len(result.state["tickets"]))        # list[Ticket]
print(result.cost_usd)                     # e.g. 0.43 (0.0 with Ollama)

Inspect the trace after any run:

antcrew inspect <run-id>
# Shows: prompt, response, tokens, cost, governance hash — per agent
antcrew trace replay <run-id>
# Replays every agent call to detect model drift

Teams

from antcrew import (
    DevTeam, FullStackTeam, ResearchTeam, ContentTeam, CustomTeam, Router,
    LegalReviewTeam, CodeMigrationTeam, ReproducibleResearchPipeline,
    BrandVoiceContentTeam, BrandVoiceProfile, WhiteLabelWrapper,
)

Features

• LLM-agnostic. Anthropic, OpenAI, Gemini, Groq, Azure, Ollama, LM Studio, LiteLLM (100+ providers). Mix models per agent.
• Local-first. Run entirely on your machine with Ollama — no API keys, no data leaves your network.
• Typed artifacts. PRDs, tickets, code, tests, and docs are Pydantic objects — predictable, auditable, and serializable.
• TraceLog. Every agent call is written to a local SQLite database. antcrew inspect <id> and antcrew trace replay <id> work offline.
• Governance hash. Each agent configuration produces a deterministic SHA-256 hash — cite in papers, pin in CI.
• Human-in-the-loop. FlexibleHITL pauses any checkpoint for local approval (callback) or remote review (via antcrew-platform).
• Semantic memory. ChromaDB or in-memory vector store. Agents reference decisions from past runs.
• EvalSuite. Regression testing for agent outputs. Run in CI with antcrew eval.
• MCP tools. Any MCP-compatible tool server works out of the box.
• Project sessions. Tickets, code, and docs accumulate across multiple runs instead of starting fresh.
• Real sandbox execution. Generated tests run in a subprocess or Docker container and results feed back into state.
• Retry + resilience. Exponential-backoff retry on timeouts, rate limits, and transient errors.

Specialized teams

Legal review:

from antcrew import LegalReviewTeam

team = LegalReviewTeam()
result = team.run(nda_text)
finding = result.state["legal_finding"]
print(f"High-risk clauses: {finding.high_risk_count}, approved: {finding.approved}")

Reproducible research — cite and replay:

from antcrew import ReproducibleResearchPipeline

pipeline = ReproducibleResearchPipeline(db_path="experiments.db")
exp = pipeline.run("What are the failure modes of multi-agent AI systems?")
print(exp.experiment_id)   # "<team_hash>:<run_id>" — stable identifier

# Replay later to detect model drift
for call in pipeline.replay_experiment(exp.experiment_id):
    print(call["agent_name"], "matched:", call["matched"])

Brand voice content:

from antcrew import BrandVoiceContentTeam, BrandVoiceProfile

profile = BrandVoiceProfile(
    name="Acme Corp",
    tone="Professional but approachable",
    standards=["Always end with a CTA", "Use 'you' not 'users'"],
    examples=["Our API ships same-day — because waiting is so 2019."],
)
team = BrandVoiceContentTeam(brand=profile)   # requires pip install antcrew[memory]
result = team.run("Write a product launch announcement")

White-label billing:

from antcrew import DevTeam, WhiteLabelWrapper

billing = WhiteLabelWrapper(DevTeam(), client_label="acme-corp", markup_pct=200)
record = billing.run("Build a REST API for a todo app")
print(f"Billed: ${record.billed_usd:.4f}  Margin: {record.margin_pct:.1f}%")

CLI reference

antcrew run "goal"              # run a team locally
antcrew run "goal" --model ollama:llama3   # offline, no API key
antcrew init                    # scaffold a new project interactively
antcrew inspect <run-id>        # view trace: prompts, tokens, cost, governance hash
antcrew trace replay <run-id>   # replay all agent calls
antcrew eval                    # run EvalSuite regression tests
antcrew describe                # show pipeline data flow (consumes/produces)
antcrew serve                   # local web dashboard
antcrew cost                    # usage and cost summary
antcrew dag                     # visualize agent graph

Run antcrew --help for the full list of 29 commands.

Optional extras

pip install "antcrew[memory]"    # ChromaDB semantic memory
pip install "antcrew[litellm]"   # 100+ LLM providers via LiteLLM
pip install "antcrew[slack]"     # Slack HITL + notifications
pip install "antcrew[telegram]"  # Telegram notifications
pip install "antcrew[mcp]"       # MCP tool servers

Architecture

antcrew is two packages shipped together:

pip install antcrew installs both. You don't need to install or import antcrew-engine directly — all its capabilities (Architect, CodeGenerator, TestRunner…) are re-exported from antcrew. The separation exists so antcrew-engine can be used standalone without LangGraph — useful if you're building a custom execution layer on top.

When to use antcrew-platform (optional)

The SDK runs entirely locally — no cloud account required. antcrew-platform is the optional SaaS layer for teams that need:

• Multi-workspace concurrent runs with cost roll-up
• Remote HITL — reviewers approve from Slack or a web link, not a terminal
• Dashboard — live run stream, eval trends, cost charts
• GitHub App — auto-post explainability comments on PRs
• Webhook delivery — push run events to your systems

→ antcrew-platform

License

MIT