fleet-deepagents-export

Runtime support for agents exported from LangSmith Fleet. Reads the export config and wires up a deepagent with MCP tools, subagents, and skills.

Published on PyPI as fleet-deepagents-export:

pip install fleet-deepagents-export

This package is usually consumed via examples/template-agent/ — a drop-in starter project that depends on it. Copy that directory, drop your Fleet export into fleet/, and run.

Quickstart

Requires uv and Python 3.11+ (uv will fetch a compatible Python if you don't have one).

Working with a coding agent (Claude Code, Cursor, etc.)? Point it at this repo and your .zip and ask it to set up the starter — the IDs it needs for .env are all in fleet/config.json after unzipping.

1. Grab the starter (it pulls fleet-deepagents-export from PyPI as a normal dependency):

   git clone --depth 1 https://github.com/langchain-ai/fleet-deepagents-export.git
   cp -R fleet-deepagents-export/examples/template-agent my-agent && cd my-agent
   

   Or, if you'd rather wire things up by hand, pip install fleet-deepagents-export into your own project and follow the layout described under What the package reads.
2. Export your agent from LangSmith Fleet (the .zip), then drop the contents into fleet/ (which ships empty apart from a placeholder README):

   unzip path/to/my-export.zip -d fleet/
   

3. Fill in .env:

   cp .env.example .env   # then edit
   

   The three LANGSMITH_*_ID values live in fleet/config.json under metadata — just copy them across.
4. Run:

   make setup    # install dependencies (uv sync)
   
   # then use one of these to interact with the agent:
   make dev      # LangGraph Studio — browser UI for chatting with the agent and inspecting graph state
   make run      # terminal REPL via cli.py — text-only chat
   

Re-exporting from Fleet later? Wipe and re-unzip — the rest of your project is untouched:

rm -rf fleet && unzip path/to/my-new-export.zip -d fleet/

What's editable in the starter

The starter separates "code from Fleet" from "code you add":

• fleet/ — owned by Fleet. Ships empty (just a placeholder README); drop your export contents here (AGENTS.md, config.json, tools.json, optional subagents/ and skills/). Re-unzip a fresh export to update; nothing else is touched.
• agent.py — graph wiring. The model comes from fleet/config.json; replace the model = components.pop("model") line to override. The starter ships with langchain-anthropic, langchain-openai, and langchain-google-genai installed — for any other provider prefix (bedrock:, mistralai:, etc.), add the matching langchain-<provider> package to pyproject.toml.
• custom_tools.py — owned by you. Add code-defined tools; merged with Fleet MCP tools at runtime.
• custom_middleware.py — owned by you. Add AgentMiddleware instances for logging, filters, pre/post hooks, etc.
• custom_skills/ — owned by you. Drop <skill-name>/SKILL.md files here; layered on top of fleet/skills/ (later overrides earlier when names collide). Files are read at module load and injected into agent state — the agent never reads disk at request time.
• cli.py — terminal REPL; edit freely for a different I/O loop.
• pyproject.toml — rename the project, pin deps.

What the package reads

From the fleet/ directory:

• AGENTS.md → system prompt
• tools.json → MCP server connections (builtin / OAuth / bearer token)
• subagents/<name>/AGENTS.md + tools.json → subagent definitions
• skills/*/SKILL.md → skill instructions

MCP auth

At startup, each tool's mcp_server_url is resolved against LangSmith's MCP server registry. One round-trip, then connections are built per the registry's auth info. Three paths:

• Builtin LangSmith tools (Gmail, Calendar, builtin GitHub) — authed via your PAT.
• Static-credential servers (auth_type: "headers") — credentials come from the registry record. Requires mcp-servers:invoke permission.
• OAuth servers (auth_type: "oauth") — bearer token fetched from LangSmith's OAuth broker. A browser opens on first run for any per-user server that isn't yet authorized for your user.

Servers not in the registry are skipped with a warning.

Env vars