Leaxp

Learn -> experience — an agentic experience kit. Package and distribute agent skills as portable, deduplicated, model-agnostic .xp files.

Leaxp learns agent skills from examples (text, code, images) and packages them into *.xp files that any agent can load and recall via MCP. Smaller context, cheaper inference, model-agnostic.

Status: pre-alpha bootstrap.

Install

# minimal
pip install leaxp

# with the local-first stack (Chroma + sentence-transformers)
pip install "leaxp[chroma,local-embed]"

# everything
pip install "leaxp[all]"

Quickstart

exp init                                    # configure providers + create leaxp.toml
exp learn ./docs --goal "use leaxp"         # extract records, embed, store
exp try ./docs/index.md --goal "use leaxp"  # debug a single doc
exp forge                                   # cluster, consolidate, dedupe (optional)
exp build my-skill.xp                       # export to a portable .xp
exp load my-skill.xp                        # load + locally re-embed
exp mcp --port 8822                         # serve via MCP for any agent
exp use --local 8822 --claude               # wire MCP into Claude Code

Concepts

• Claim — undisputable fact the agent must know but cannot modify.
• Hint — adjustable best-practice the agent can adapt to context.
• Recall — the MCP tool agents call; returns records grouped by Useful WHEN: <condition>.
• .xp file — zip with a manifest + JSONL of raw records (no embeddings). Embeddings are computed locally on exp load so the same pack works with any embedding model.

Project structure

See CLAUDE.md for the package architecture overview.

Development

uv sync                                  # create venv, install deps
uv run pre-commit install                # enable hooks
uv run ruff check . && uv run pyright    # lint + typecheck
uv run pytest --cov                      # tests with coverage
uv run exp --help

License

Apache-2.0 — see LICENSE.txt.