generated: '2026-08-27'
method: searched
source: https://localai.io/features/localai-assistant
docs: https://localai.io/features/localai-assistant
name: LocalAI Assistant MCP server
status: published
deployment:
  mode: local-stdio
  endpoint: null
  install: 'local-ai mcp-server --target http://remote.localai:8080 --api-key <admin-key>'
  package: https://github.com/mudler/LocalAI
  auth: api-key
  verified: searched
deployment_note: >-
  There is no vendor-hosted remote MCP endpoint. The provider ships the server two ways, both requiring a
  LocalAI instance the operator runs. (1) In-process: when an admin flips the Manage toggle on a chat
  session, the conversation is wired to an in-process MCP server — no loopback socket, no synthetic API
  key, no extra TCP port — so it is reachable only from inside that LocalAI process, never by an external
  agent. (2) Standalone stdio: `local-ai mcp-server --target <url> --api-key <admin-key>` runs the same
  tool catalog as a stdio server pointed at any LocalAI HTTP API, for hooking into Claude Desktop, Cursor
  or another MCP host. `--read-only` skips registration of every mutating tool. The tool catalog is
  identical in both variants. The mode is therefore local-stdio, not remote: a human must install and run
  it before any agent can reach it.
tool_source: >-
  Verbatim from the published tool catalog in the LocalAI Assistant documentation, plus the middleware
  documentation's two MCP tool tables. Input schemas were not read — that requires authenticated
  introspection against a running instance, which this pass had no access to.
transport: stdio
protocol: Model Context Protocol
tools:
- name: gallery_search
  category: models
  access: read
- name: list_installed_models
  category: models
  access: read
- name: list_galleries
  category: models
  access: read
- name: list_backends
  category: backends
  access: read
- name: list_known_backends
  category: backends
  access: read
- name: get_job_status
  category: jobs
  access: read
- name: get_model_config
  category: config
  access: read
- name: vram_estimate
  category: models
  access: read
- name: system_info
  category: monitoring
  access: read
- name: list_nodes
  category: p2p
  access: read
- name: get_pii_events
  category: middleware
  access: read
  description: Recent redaction / block events with optional filters.
- name: get_middleware_status
  category: middleware
  access: read
  description: Aggregator returning the same payload as GET /api/middleware/status.
- name: get_router_decisions
  category: router
  access: read
  description: Recent router decision log with optional filters.
- name: get_router_corpus_stats
  category: router
  access: read
  description: KNN corpus size and per-label counts; entry texts are never returned.
- name: install_model
  category: models
  access: write
- name: import_model_uri
  category: models
  access: write
- name: delete_model
  category: models
  access: write
- name: install_backend
  category: backends
  access: write
- name: upgrade_backend
  category: backends
  access: write
- name: edit_model_config
  category: config
  access: write
- name: reload_models
  category: models
  access: write
- name: toggle_model_state
  category: models
  access: write
- name: toggle_model_pinned
  category: models
  access: write
- name: seed_router_corpus
  category: router
  access: write
  description: Add labelled exemplars to a KNN router's corpus.
- name: clear_router_corpus
  category: router
  access: write
  description: Wipe a KNN router's corpus.
tool_count: 25
safety:
  admin_only: true
  confirmation: >-
    Mutating tools are guarded by a system-prompt rule requiring the model to confirm the action with the
    user before calling them. There is no separate code-side preview/apply step.
  read_only_flag: --read-only
  disable: LOCALAI_DISABLE_ASSISTANT=true (chat requests with metadata.localai_assistant=true then return 503)
mcp_client_capability:
  note: >-
    Distinct from the server above, LocalAI is also an MCP HOST — it connects models and agents outward to
    remote HTTP and local stdio MCP servers and executes their tools server-side inside the completion
    loop. This is configuration, not an agent surface LocalAI exposes, so it does not affect deployment.mode.
  endpoints:
  - POST /v1/mcp/chat/completions
  - POST /mcp/v1/chat/completions
  - POST /mcp/chat/completions
  selection: metadata.mcp_servers (comma-separated server names, stripped before reaching the backend)
  scopes:
  - model-scoped via the `mcp:` block in a model YAML (remote + stdio)
  - agent-scoped via an agent's mcp_servers / mcp_stdio_servers fields
  features:
  - MCP Prompts — discover and expand reusable prompt templates from MCP servers
  - MCP Resources — browse and inject resource content from MCP servers into conversations
  - Bearer token authentication to remote MCP servers
  - agent.max_iterations caps the tool execution loop (default 10)
  docs: https://localai.io/features/mcp