Pyfa MCP

uvx pyfa-mcp

git clone --recurse-submodules git@github.com:theonlysinjin/eve-fit-mcp.git
cd eve-fit-mcp
uvx --from . pyfa-mcp

uvx --from git+https://github.com/theonlysinjin/eve-fit-mcp.git pyfa-mcp

{
  "mcpServers": {
    "pyfa-mcp": {
      "command": "uvx",
      "args": ["--from", "/path/to/eve-fit-mcp", "pyfa-mcp"]
    }
  }
}

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

git clone --recurse-submodules git@github.com:theonlysinjin/eve-fit-mcp.git
cd eve-fit-mcp
uv venv .venv && source .venv/bin/activate
uv pip install -e ".[dev]"
pytest
pyfa-mcp   # or: python -m pyfa_mcp

# Fit with Pyfa MCP

You design EVE Online fits by proposing changes; the **pyfa-mcp** MCP evaluates them. It never suggests modules — you do.

Use the **eve-online-esi** MCP when the fit should reflect a real character’s skills. Prefer a cached map under `users/<Name>/skills.json`; refresh from ESI when asked or when the file is missing/stale.

## Skills (player maps)

**Apply (fitting):** load `users/<Name>/skills.json` → `set_skills(fit_id, data["skills"])`. Keys are skill type IDs; values are `active_skill_level` (0–5). Theorycraft: skip and use `apply_all_skills_5`.

**Refresh (update cache):**
1. `add_character` if needed (SSO; tokens stay local). Re-auth on 401 / missing character.
2. `GetCharactersCharacterIdSkills` with `character_id` (+ `X-Compatibility-Date`).
3. Write `users/<Name>/skills.json` as `{ character_id, name, updated_at, total_sp?, unallocated_sp?, skills: { str(skill_id): active_skill_level } }`.

## Startup

1. Confirm the goal in one line: role, constraints (EHP, DPS, tank type, cap stable?), and skills (player map via `users/…/skills.json`, or theorycraft).
2. If using a player map: load (or refresh) skills as above.
3. `create_fit(ship_type_id)` then either `set_skills` (from file) or `apply_all_skills_5`.
4. Rough in a full fit with type IDs: highs → mids → lows → rigs → drones/fighters → implants if needed.
5. Read the FitReport. Fix hard blockers first (`validation_errors`, CPU/PG/slots), then optimize toward the goal.
6. Iterate: **one** change per turn (`equip_module` / `replace_module` / `set_module_state` / `set_charge` / …). Compare reports. Use `clone_fit` for A/B forks.

## Rules

- Type IDs only — never invent them.
- Soft failures (CPU, skills, slots) still apply; hard errors (bad ID, wrong rack) do not mutate.
- Racks: `high` / `mid` / `low`. States: `offline` | `online` | `active` | `overload`.
- Stop when constraints are met or gains flatten. Summarize the final fit + key stats.

## FitReport priorities

`validation_errors` → `resources` / `slots` → `combat` (dps, ehp, RPS) → `mobility` → `fit` snapshot.