generated: '2026-08-27'
method: generated
source: >-
  openapi/localai-api-openapi.yml plus the LocalAI documentation (authentication, api-errors, mcp,
  middleware, tracing, localai-assistant, api-discovery)
api: LocalAI API
note: >-
  Four packaged Agent Skills covering the flows a LocalAI instance is actually driven through. Every
  operation named in every skill was verified present in the published Swagger document before it was
  written; the handful that are documented but absent from the contract (/api/pii/events,
  /api/middleware/status, /api/router/decisions, /api/operations/{id}/cancel) are called out as
  unverified in the skill text rather than presented as contract operations. Because the contract declares
  no operationId on any of its 123 operations, operations are referenced as "METHOD path" throughout —
  that is the only stable identifier this contract offers.
  Searched first for provider-published agent instructions and FOUND them: LocalAI publishes an
  AGENTS.md at the repository root indexing 16 topic guides under .agents/, saved verbatim below.
  Separately, LocalAI's Skills system (LOCALAI_AGENT_POOL_ENABLE_SKILLS, /api/agents/skills, git-based
  skill repositories) is a runtime feature for skills the OPERATOR authors, not a catalogue of published
  skills for driving the API, so there was nothing to save from it.
skills:
- name: localai-orient-against-an-instance
  file: localai-orient-against-an-instance.md
  summary: >-
    Establish what an unknown LocalAI instance is, which subsystems are live, which models it serves and
    what the caller is allowed to call — using the three anonymous discovery endpoints first.
  operations: 6
- name: localai-run-an-inference-request
  file: localai-run-an-inference-request.md
  summary: >-
    Choose the right compatibility surface (OpenAI, Anthropic, Open Responses, ElevenLabs), send the
    request, and parse errors with the envelope that endpoint family actually uses.
  operations: 8
- name: localai-install-a-model
  file: localai-install-a-model.md
  summary: >-
    Search the gallery, rehearse the VRAM fit, confirm with the human, install, poll the job, verify —
    and know which of the surrounding writes have no documented inverse.
  operations: 8
- name: localai-operate-and-observe
  file: localai-operate-and-observe.md
  summary: >-
    Diagnose a failed or rewritten request through traces, backend logs, middleware decision and event
    logs, and account for token usage per user and per API key.
  operations: 9
provider_published:
- name: LocalAI Agent Instructions
  file: localai-AGENTS.md
  method: searched
  source: https://raw.githubusercontent.com/mudler/LocalAI/master/AGENTS.md
  http_status: 200
  audience: coding-agents-contributing-to-localai
  note: >-
    Saved verbatim. This is an entry point for AI coding assistants working ON LocalAI (Claude Code,
    Cursor, Copilot, Codex, Aider), not for agents CALLING the API — a distinction worth keeping straight.
    It is unusually substantive for the genre: LocalAI adopts the Linux kernel project's guidelines for AI
    coding assistants, forbids Signed-off-by and Co-Authored-By trailers from an AI, and mandates an
    `Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2]` trailer instead, with the human submitter
    owning the change. Two of its rules are directly relevant to this catalog: a docs-with-code rule
    requiring docs/content/ to be updated in the same change as any user-facing API/CLI/config change,
    and a rule that every admin endpoint an admin would manage conversationally MUST also be exposed as an
    MCP tool in pkg/mcp/localaitools/, enforced by a TestToolHTTPRouteMappingComplete test. The second is
    the governance behind mcp/localai-tool-crosswalk.yml — REST/MCP drift is treated as a test failure.
  topic_guides: 16
  topic_guide_index: https://github.com/mudler/LocalAI/tree/master/.agents