API overview: KoboldAI, OpenAI and Ollama APIs

KoboldCpp serves several APIs at once, all on the same port (5001 by default). Many apps built for other servers, such as OpenAI, Ollama or Automatic1111, can use KoboldCpp once you point them to its address. Not every feature of those servers is supported.

API families

API	Base path	Use it for
KoboldAI (native)	`/api/v1/`, `/api/extra/`	Text generation with full sampler control, plus KoboldCpp extras such as token counting, abort and multiplayer. KoboldAI Lite uses this API.
OpenAI	`/v1/`	Completions, chat completions (with tools and images), embeddings, audio and images
OpenAI Responses	`/v1/responses`	Apps built for the OpenAI Responses API
Anthropic Messages	`/v1/messages`	Apps built for the Anthropic API, with images and tool calls
Ollama	`/api/generate`, `/api/chat`, `/api/embed`, `/api/tags`	Ollama apps; run KoboldCpp on port 11434 so they find it
A1111 / Forge	`/sdapi/v1/`	Image generation, upscaling and image captioning
ComfyUI	`/prompt`, `/view`, `/history`	Apps that send image jobs to ComfyUI
XTTS	`/tts_to_audio`, `/speakers_list`	Apps built for an XTTS text-to-speech server
OpenAI audio (Whisper)	`/v1/audio/transcriptions`, `/v1/audio/speech`	Speech-to-text and text-to-speech

Each family needs the matching model loaded. Text endpoints need a text model, image endpoints an image model, and so on. Endpoints lists every path and what it needs.

API documentation

On your instance: open http://localhost:5001/api for interactive API docs.
Online: lite.koboldai.net/koboldcpp_api has the same docs.

Both cover the KoboldAI, OpenAI, Anthropic and A1111 endpoints. The Ollama, ComfyUI and XTTS emulation is not in them; see Endpoints.

First requests

Two examples with curl. They use Linux and macOS shell quoting.

KoboldAI API:

curl http://localhost:5001/api/v1/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Once upon a time", "max_length": 50}'

The reply text is in results[0].text.

OpenAI chat completions:

curl http://localhost:5001/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Hello!"}], "max_tokens": 50}'

On Windows, Command Prompt needs the command on one line, with the inner quotes escaped. In PowerShell, type curl.exe instead of curl.

curl http://localhost:5001/api/v1/generate -H "Content-Type: application/json" -d "{\"prompt\": \"Once upon a time\", \"max_length\": 50}"

If you set a password, add -H "Authorization: Bearer <password>". See Passwords and security.

What is loaded

Path	Shows
`GET /api/extra/version`	Which features are available: `llm`, `txt2img`, `vision`, `audio`, `transcribe`, `tts`, `embeddings`, `music`, `websearch`, `multiplayer`, `jinja`, `mcp`, `admin`, `router`, `protected` and more
`GET /v1/models`	The loaded text model; in router mode also the files you can switch to
`GET /props`	Chat template, context size (`n_ctx`) and whether vision and audio input work
`GET /api/extra/true_max_context_length`	The context size the server allocated
`GET /api/extra/perf`	Speed of the last request and the queue length

Chat formatting

The chat-style APIs (OpenAI chat, Responses, Anthropic, Ollama chat) send messages, and KoboldCpp turns them into the prompt format the model expects.

By default, KoboldCpp guesses the format from the model's chat template (Chat Adapter: on the Loaded Files tab, --chatcompletionsadapter, default AutoGuess). The console prints Chat completion heuristic: <name>.
Use Jinja (--jinja) renders the model's own chat template instead. Try it if chat replies look badly formatted with a newer model.
For tool calls with Jinja, also tick Jinja for Tools (--jinjatools).

Advanced: server-side defaults

These apply to every request from every app:

Launcher (Context tab)	Flag	Effect
Default Gen Amt:	`--defaultgenamt`	Reply length when the request sets none (default 2048, at most half the context)
Prompt Limit:	`--genlimit`	Hard cap on reply length; longer requests are cut to it
Default Params:	`--gendefaults`	JSON with default values for fields the request leaves out, using the native KoboldAI and A1111 field names, e.g. `{"rep_pen": 1.1}`. For the reply length, use Default Gen Amt: instead.
Override	`--gendefaultsoverwrite`	Makes Default Params: overwrite values the request did send

Advanced: routing details

Most paths are matched by their ending; a few, such as /props and /prompt, must match exactly. The query string and a trailing slash don't change which endpoint answers: http://localhost:5001/api/v1/generate/ works the same as without the slash.
Paths starting with /lcpp/ are mapped to the root, so /lcpp/v1/models answers like /v1/models. This is how the bundled llama.cpp WebUI at /lcpp/ reaches the API.
An unknown GET path returns 404 with a short HTML page that links to /api. An unknown POST path returns 404.
POST /mcp is a JSON-RPC proxy to the MCP servers from MCP JSON: (--mcpfile). See Web search and MCP.
Admin mode adds /api/admin/ endpoints for swapping models. See Admin mode.