MCP
One endpoint, a curated catalog
An MCP endpoint at /mcp exposes document and knowledge tools to assistants. Which tools are reachable is an operator decision, not a client one.
LM-Kit One · Compatibility
The coverage matrix, stated.
What each API dialect serves, endpoint by endpoint, including what is missing. A stated limitation beats a discovered one.
OpenAI dialect.
Deep enough that OpenAI-shaped retrieval and agentic flows run against a private deployment, not only single chat turns.
| Endpoint | Status | Notes |
|---|---|---|
| Inference | ||
POST /v1/chat/completions |
Served | Streaming, tools, vision content, structured output; parameters below |
POST /v1/completions |
Served | Legacy completions, with per-token logprobs and alternatives on request |
POST /v1/embeddings |
Served | Any embedding-capable model in the catalog |
GET /v1/models, GET /v1/models/{id} |
Served | The list envelope OpenAI SDKs parse; the Anthropic shape when the request carries anthropic-version |
| Files and retrieval | ||
/v1/files upload, list, retrieve, delete |
Served | Scoped to the calling key; purpose is recorded and filters the list |
/v1/vector_stores create, list, get, delete |
Served | Backed by the built-in search engine |
/v1/vector_stores/{id}/files attach, list, retrieve, detach |
Served | Status moves from in_progress to completed as the embedding pass lands, so SDK create-and-poll helpers work |
| Responses API | ||
POST /v1/responses |
Served | Create; retrieve, delete and input items below |
GET /v1/responses/{id}, GET .../input_items, DELETE |
Served | Stored responses round-trip |
POST /v1/responses/{id}/cancel |
Not served | Background execution is refused by name, so there is nothing to cancel |
Chat parameters: honored, extended, or ignored.
Compatibility claims usually die on parameters, so here is the split for
/v1/chat/completions.
| Parameters | Status | Notes |
|---|---|---|
temperature, top_p, max_tokens / max_completion_tokens, stop, seed, frequency_penalty, presence_penalty, logit_bias |
Honored | Mapped onto the engine's sampling controls |
stream, stream_options |
Honored | Server-sent events, with usage in the final chunk |
response_format (json_object, json_schema) |
Enforced | Grammar-constrained decoding: the completion parses, or the request is refused up front with an OpenAI-shaped error |
tools, tool_choice |
Honored | Exercised end to end by real agentic clients |
logprobs, top_logprobs |
Honored | Per-token log probabilities with up to 20 alternatives per position |
agent, skill, server_tools, memory, reasoning_effort, thinking |
LM-Kit extensions | Adopt a named server-side agent, pin a skill, enable governed server tools and persistent memory, set thinking effort |
Legacy functions, function_call |
Refused by name | A 400 points at tools; the deprecated shape is never approximated |
parallel_tool_calls, and wire-compatibility fields such as service_tier, audio, prediction |
Accepted, ignored | Parsed so clients do not break; no engine equivalent yet |
Anthropic dialect.
The Messages shape, so SDKs and coding agents built on it keep working while inference stays inside your network.
| Endpoint | Status | Notes |
|---|---|---|
POST /v1/messages |
Served | Streaming, system, tools and tool_choice, stop_sequences, temperature / top_p / top_k, and thinking with a token budget |
GET /v1/models, GET /v1/models/{id} |
Served | The Anthropic shape, paged, selected by the anthropic-version header every SDK sends |
POST /v1/messages/count_tokens |
Served | Exact from the model's own tokenizer when the model is resident, estimated otherwise: a count never costs a model load |
Ollama dialect.
The whole surface, lifecycle included, which is the part shims normally skip: the stock CLI drives this server unmodified.
| Endpoint | Status | Notes |
|---|---|---|
/api/chat, /api/generate |
Served | Streaming and non-streaming |
/api/embed, /api/embeddings |
Served | Current and legacy embedding shapes |
/api/tags, /api/show, /api/ps, /api/version |
Served | Discovery and status |
/api/pull, /api/create, /api/copy, /api/delete |
Served | Model lifecycle against this server's catalog and store |
/api/blobs/{digest} (HEAD, POST) |
Served | A multi-gigabyte model upload is a tested path |
/api/push |
Not served | By design: this server is not a model registry, and the endpoint says so instead of pretending |
MCP and the native surface.
The dialects are the on-ramp. The native REST API is the whole server, and MCP is how assistants reach a governed slice of it.
Native REST
Over a hundred endpoints
Documents, extraction, search, agents, PDF operations, training and model management: operations no general chat API models. The server ships its own API reference.
Anything not listed on this page is not served. The matrix is maintained with each release; if a client you depend on needs an endpoint marked missing, tell us.
LM-Kit One