Flyte

flyte-mcp

Flyte V2 knowledge, patterns, plugins, and runtime — exposed to every AI coding assistant via the Model Context Protocol.

Claude Code using flyte-mcp

Install in one click

Claude Code

claude mcp add flyte -- uvx flyte-mcp

Claude Desktop / any MCP client

{
  "mcpServers": {
    "flyte": {
      "command": "uvx",
      "args": ["flyte-mcp"]
    }
  }
}

Add this to ~/.claude.json, ~/.cursor/mcp.json, or your client's config file.

Why it exists

Ask any AI assistant "write a Flyte V2 task with caching and 4 GPUs" and you get confidently wrong code: V1 imports, invented decorators, hallucinated resource kwargs. The assistant has no reliable channel into the Flyte ecosystem, so it fills the gap with training-data guesses.

flyte-mcp is that channel. It ships a versioned knowledge pack built directly from the flyte-sdk source tree and the Flyte Plugin Registry, plus a thin runtime bridge for executing tasks when a cluster is configured. The assistant stops guessing and starts answering.

What your assistant can do

Capability	Tools
Learn the V2 API	`get_flyte_symbol` · `search_flyte_api` · `list_flyte_symbols`
Find canonical examples	`find_flyte_example_for` · `get_flyte_pattern` · `list_flyte_patterns`
Pick the right plugin	`suggest_flyte_plugin_for` · `list_flyte_plugins` · `get_flyte_plugin`
Port V1 code to V2	`migrate_v1_to_v2`
Get oriented	`get_flyte_overview` · `get_flyte_features` · `get_flyte_install_guide` · `get_flyte_version`
Run on a cluster	`run_flyte_task` · `get_flyte_execution_status` · `list_flyte_recent_runs`

All tools are pure Python, stdio transport, zero network calls unless you explicitly use the runtime bridge.

Example prompts that just work

How do I cache a task and invalidate on input change?
Show me a distributed PyTorch training example with A100s.
Which Flyte plugin do I use for Snowflake, and what's the import?
Migrate this flytekit V1 workflow to V2.
What's the signature of TaskEnvironment?

Your assistant picks the right tools and assembles accurate answers.

Rebuilding the knowledge pack

Contributors and release automation can regenerate the pack from source:

python scripts/build_knowledge.py \
  --sdk-path /path/to/flyte-sdk \
  --registry /path/to/flyte-plugin-registry/src/data/plugins.json \
  --out src/flyte_mcp/data/flyte-v2-knowledge.json

Sources used:

flyte-sdk/src/flyte/__init__.py — public symbols via AST
flyte-sdk/examples/* — canonical patterns by theme
flyte-sdk/README.md, FEATURES.md, CONTRIBUTING.md — meta docs
flyte-plugin-registry — curated plugin catalog

flytesnacks is intentionally excluded: V2 consolidated examples in-tree.

Relationship to other Flyte MCP projects

wherobots/flyte-mcp — runtime-only. Discovers and executes tasks on a deployed Flyte instance via API key. Complementary, not competing.
unionai/claude-agents-public — Claude Code custom agents (system prompts, not an MCP server). Compose freely.

This project focuses on authoring: the moment a developer types a prompt asking about Flyte.

Roadmap

GitHub Action to auto-rebuild the knowledge pack on every flyte-sdk release
Listing in the official MCP Registry
Local semantic search via small sentence-transformer model
AST-based migration codemod (replacing the current regex pass)
Log streaming tool (get_flyte_execution_logs) with tail support

License

Apache-2.0 — same license as Flyte itself.

Disclaimer

Independent community project. Not officially affiliated with or endorsed by Flyte or Union.ai. The Flyte name and logo are trademarks of their respective owners.

flyte-mcp

flyte-mcp

Install in one click

Why it exists

What your assistant can do

Example prompts that just work

Rebuilding the knowledge pack

Relationship to other Flyte MCP projects

Roadmap

License

Disclaimer

Reviews