langgraph-forge 0.4.0a1

Initialise LangGraph-based agent architectures: pick a provider, a pattern, a deployment — get a runnable project.

langgraph-forge

Initialise LangGraph-based agent architectures: pick a provider, a pattern, a deployment — get a runnable project.

Why?

LangGraph's primitives — init_chat_model, create_react_agent, langgraph-supervisor, langchain-mcp-adapters, cloud-runtime wrappers — are individually excellent but scattered across five packages with inconsistent config shapes. Every team building an agent re-writes the same wiring.

langgraph-forge absorbs that wiring into thin, opinionated factories plus a CLI scaffolder. We compose, we don't reimplement. The value is a coherent surface and a trade-off-aware starter, not new primitives.

Install

pip install langgraph-forge
# or, for optional cloud adapters:
pip install 'langgraph-forge[bedrock,vertex,azure]'

Quickstart (under a minute)

# 1. Scaffold a project
langgraph-forge init my-agent --provider anthropic --pattern supervisor --deploy direct --no-input
cd my-agent

# 2. Install + set credentials
cp .env.example .env        # then edit
uv sync --extra dev

# 3. Smoke-test the generated graph (uses mocked LLM)
uv run pytest tests/unit -q

# 4. Run it
uv run python -m my_agent

Matrix

Provider (--provider)	Pattern (--pattern)	Deployment (--deploy)
anthropic, openai, grok, google, bedrock, azure	single, supervisor	direct, bedrock, vertex, azure

All combinations scaffold end-to-end. direct is fully functional in v0.1; cloud adapters (bedrock / vertex / azure) ship as Protocol-conformant contract stubs — the scaffolded deploy.py imports successfully and the smoke test passes, but calling the cloud prepare / invoke raises NotImplementedError until the SDK glue lands in v0.2.

Library usage (without the CLI)

from langgraph_forge import (
    ModelSpec,
    SpecialistSpec,
    create_supervisor_agent,
    get_model,
)

supervisor = get_model(ModelSpec(model="claude-opus-4-7", provider="anthropic"))
worker_model = ModelSpec(model="claude-haiku-4-5", provider="anthropic")

graph = create_supervisor_agent(
    supervisor_model=supervisor,
    specialists=[
        SpecialistSpec(
            name="researcher",
            prompt="You gather facts.",
            model=worker_model,
        ),
        SpecialistSpec(
            name="summariser",
            prompt="You produce concise summaries.",
            model=worker_model,
        ),
    ],
    supervisor_prompt="Delegate research and summarisation to specialists.",
)

Swap providers by changing provider="anthropic" to "openai", "xai", "google_genai", "bedrock_converse", or "azure_openai". Swap deployment by replacing DirectAdapter with BedrockAgentCoreAdapter / VertexAgentEngineAdapter / AzureAIAgentAdapter.

MCP integration

from langgraph_forge import MCPConfig, MCPServerConfig, load_mcp_tools

config = MCPConfig(
    servers={
        "filesystem": MCPServerConfig(
            transport="stdio",
            command="npx",
            args=["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
        ),
    }
)
tools = await load_mcp_tools(config)

Extending: writing a new deployment adapter

Every adapter satisfies the DeploymentAdapter Protocol (name, extras, prepare, invoke, template fragment). Third-party packages register via the langgraph_forge.deployment_adapters entry-point group, which means adding a fifth target does not require a PR to this repo.

# your_package/pyproject.toml
[project.entry-points."langgraph_forge.deployment_adapters"]
modal = "your_package.modal_adapter:ModalAdapter"

After pip install your-package, langgraph-forge list-deploy includes modal and --deploy modal works in init.

Not included — by design

Listed so expectations stay calibrated. These are out of scope in v0.1, many permanently:

• Swarm pattern / non-supervisor multi-agent topologies (may return in v0.3+)
• LangSmith / OpenTelemetry / tracing / metrics
• Prompt versioning, PromptPack, content_hash, prompt registry, eval harness
• Tool allowlists, SideEffectGate, autonomy gates, budget / cost ceilings, cycle detection
• Schema registry, output-envelope validation, per-specialist output_type
• Peer review, human review queue, ExceptionTicket, routing hints
• HTTP / REST serving, WebSocket, SSE, web UI
• Auth(n/z), permissions, multi-tenancy, rate limiting
• DynamoDB / S3 / custom persistence beyond langgraph.checkpoint.*
• Model-chain fallback (compose LangChain's with_fallbacks yourself)
• Streaming helpers (graph.astream() is already the answer)

Documentation

Deeper material lives under docs/, organised by audience and (for user docs) by Diátaxis quadrant:

• Tutorials — learn by doing
• How-to guides — task recipes
• Reference — exhaustive lookup (start with CLI reference)
• Explanation — concepts and design
• Contributing — developer docs (architecture, testing, ADRs)
• Governance — maintainer ops

Versioning

Pre-1.0; see VERSIONING.md for the full policy (breaking-change rule, deprecation grace period, yank policy, and the gates to 1.0). Release notes live in GitHub Releases.

Contributing

See CONTRIBUTING.md for the full flow (fork → branch → PR → review). Deeper material in docs/contributing/.

Security

See SECURITY.md for private disclosure.

License

MIT — see LICENSE.