llama-swap

Run multiple generative AI models on your machine and hot-swap between them on demand. llama-swap works with any OpenAI and Anthropic API compatible server and is used by thousands of people to power their local AI workflows.

Built in Go for performance and simplicity, llama-swap has zero dependencies and is incredibly easy to set up. Get started in minutes - just one binary and one configuration file.

Features:

• ✅ Easy to deploy and configure: one binary, one configuration file. no external dependencies
• ✅ On-demand model switching for many local AI servers (llama.cpp + forks, vllm, stable-diffusion.cpp, audio.cpp, ComfyUI, etc.)
• future proof, upgrade your inference servers at any time.
• ✅ OpenAI API supported endpoints:
• v1/completions
• v1/chat/completions
• v1/responses
• v1/embeddings
• v1/models - list available models
• v1/audio/speech (#36)
• v1/audio/transcriptions (docs)
• v1/audio/voices
• v1/images/generations
• v1/images/edits
• ✅ Anthropic API supported endpoints:
• v1/messages
• v1/messages/count_tokens
• ✅ llama-server (llama.cpp) supported endpoints
• v1/rerank, v1/reranking, /rerank
• /infill - for code infilling
• /completion - for completion endpoint
• /models - list available models. same behavior as v1/models
• /props - requires ?model={model_id} query parameter to be provided. The autoload parameter is not supported and will be ignored.
• ✅ SDAPI via stable-diffusion.cpp's server
• /sdapi/v1/txt2img
• /sdapi/v1/img2img
• /sdapi/v1/loras - requires model in request body to fetch the correct loras
• ✅ audio.cpp supported extra endpoints
• /audioapi/v1/tasks/run
• ✅ /audiocpp/ - audio.cpp UI endpoint, served by the audiocpp_auto model (#1221)
• ✅ /comfyui/ - ComfyUI custom endpoint (#1001) for more reliable swapping
• ✅ llama-swap API
• /ui - web UI
• /upstream/:model_id - direct access to upstream server (demo)
• /running - list currently running models (#61)
• POST /api/models/unload - manually unload all running models (#58)
• POST /api/models/unload/:model_id - unload a specific model
• POST /api/models/pause - unload all models and reject inference requests until resumed (#1268)
• POST /api/models/resume - accept inference requests again
• GET /api/profiles - list configured profiles and the active selection
• PUT /api/profiles/active - activate a profile or select none
• /logs - remote log monitoring
  • GET /logs returns buffered plain text logs.
  • If Accept: text/html is sent, /logs redirects to /ui/.
  • GET /logs/stream keeps the connection open for live log streaming.
  • Stream endpoints send buffered history first by default; add ?no-history to stream only new lines.
  • GET /logs/stream/proxy streams proxy logs only.
  • GET /logs/stream/upstream streams upstream process logs only.
  • GET /logs/stream/http streams the HTTP access log only.
  • GET /logs/stream/{model_id} streams logs for one model (including IDs with slashes, like author/model).
• /health - just returns "OK"
• /metrics - system and GPU metrics for prometheus
• ✅ API Key support - define keys to restrict access to API endpoints
• ✅ Customization
• Switch model ID routing at runtime with profiles
• Run concurrent models with a custom DSL swap matrix (#643)
• Automatic unloading of models after timeout by setting a ttl
• Docker and Podman support using cmd and cmdStop together
• Preload models on startup with hooks (#235)
• Apply filters to requests to control inference with stripParams, setParams and setParamsByID

Web UI

llama-swap includes a real time web interface with a playground for testing out all sorts of local models:

View detailed token metrics:

Inspect request and responses:

Manually load and unload models:

Real time log streaming:

Installation

llama-swap can be installed in multiple ways

1. Docker
2. Homebrew (macOS and Linux)
3. MacPorts (macOS)
4. WinGet
5. From release binaries
6. From source

Docker Install (download images)

Two types of container images are built nightly for llama-swap:

1. A unified container with llama-server, ik-llama-server, stable-diffusion.cpp, whisper.cpp, audio.cpp and llama-swap all built from source. Available for CUDA 12, CUDA 13 and Vulkan. This one is recommended for use.
2. A legacy image, which is llama.cpp's own llama-server container with llama-swap copied in. It carries only what that base image ships, so no image generation, speech or ik-llama-server.

Unified container (Recommended)

There are three unified images. Pick the one that matches your GPU:

unified-cuda13 is a multi-arch tag, so docker pull picks the right image for the host — including the aarch64 one a DGX Spark needs.

$ docker pull ghcr.io/mostlygeek/llama-swap:unified-cuda13

# run with a custom configuration and models directory
$ docker run -it --rm --runtime nvidia -p 9292:8080 \
 -v /path/to/models:/models \
 -v /path/to/custom/config.yaml:/etc/llama-swap/config/config.yaml \
 ghcr.io/mostlygeek/llama-swap:unified-cuda13

Configuring startup with environment variables

The unified images can be configured with LLAMA_SWAP_* environment variables instead of command line flags, which is usually easier in a compose file or a Kubernetes manifest. Each one maps to a llama-swap flag:

# configure startup with environment variables instead of flags
$ docker run -it --rm --runtime nvidia -p 9292:8080 \
 -v /path/to/models:/models \
 -e LLAMA_SWAP_CONFIG=/models/llama-swap.yaml \
 -e LLAMA_SWAP_LISTEN=0.0.0.0:8080 \
 -e LLAMA_SWAP_WATCH_CONFIG=false \
 ghcr.io/mostlygeek/llama-swap:unified-cuda13

services:
  llama-swap:
    image: ghcr.io/mostlygeek/llama-swap:unified-cuda13
    ports:
      - "9292:8080"
    volumes:
      - /path/to/models:/models
    environment:
      LLAMA_SWAP_CONFIG: /models/llama-swap.yaml
      LLAMA_SWAP_LISTEN: 0.0.0.0:8080
      LLAMA_SWAP_WATCH_CONFIG: "false"

Keep LLAMA_SWAP_LISTEN on 0.0.0.0 whenever you publish a port. A container that binds its own loopback is not reachable through -p, so localhost:8080 there gives connection refused. Bind loopback only when the container shares the host's network, where it usefully limits llama-swap to the host itself:

# reachable from the host only, not from the network
$ docker run -it --rm --runtime nvidia --network host \
 -v /path/to/models:/models \
 -e LLAMA_SWAP_LISTEN=localhost:8080 \
 ghcr.io/mostlygeek/llama-swap:unified-cuda13

An unset or empty variable contributes nothing and leaves llama-swap's own default. Booleans accept true/false, 1/0, yes/no or on/off in any case; anything else stops the container rather than being read as "off". -version is deliberately not mapped — use docker run <image> -version.

Passing flags to the container still works and behaves exactly as it always has: arguments replace every default rather than adding to the variables above.

# unchanged: runs `llama-swap -config /models/my.yaml`, nothing else added
$ docker run ghcr.io/mostlygeek/llama-swap:unified-cuda13 -config /models/my.yaml

See docker/unified/README.md for the details.

Legacy container

This image is llama.cpp's own llama-server container (ghcr.io/ggml-org/llama.cpp) with the llama-swap binary copied into it. It tracks llama.cpp's nightly server images closely, which is its only real advantage.

Prefer the unified container. The legacy image inherits whatever the llama-server base ships and nothing else, so it has no stable-diffusion.cpp, whisper.cpp, audio.cpp or ik-llama-server, and no LLAMA_SWAP_* environment variable support. Use it only if staying on llama.cpp's exact server image matters to you.

$ docker pull ghcr.io/mostlygeek/llama-swap:cuda

# run with a custom configuration and models directory
$ docker run -it --rm --runtime nvidia -p 9292:8080 \
 -v /path/to/models:/models \
 -v /path/to/custom/config.yaml:/app/config.yaml \
 ghcr.io/mostlygeek/llama-swap:cuda

# pull latest images per platform
docker pull ghcr.io/mostlygeek/llama-swap:cpu
docker pull ghcr.io/mostlygeek/llama-swap:cuda
docker pull ghcr.io/mostlygeek/llama-swap:vulkan
docker pull ghcr.io/mostlygeek/llama-swap:intel
docker pull ghcr.io/mostlygeek/llama-swap:musa

# tagged llama-swap, platform and llama-server version images
docker pull ghcr.io/mostlygeek/llama-swap:v166-cuda-b6795

# non-root cuda
docker pull ghcr.io/mostlygeek/llama-swap:cuda-non-root

Homebrew Install (macOS/Linux)

brew tap mostlygeek/llama-swap
brew install llama-swap
llama-swap --config path/to/config.yaml --listen localhost:8080

MacPorts (macOS)

[!NOTE] Maintained by MacPorts community - llama-swap port. It is not an official part of llama-swap.

sudo port install llama-swap
llama-swap --config path/to/config.yaml --listen localhost:8080

WinGet Install (Windows)

[!NOTE] WinGet is maintained by community contributor Dvd-Znf (#327). It is not an official part of llama-swap.

# install
C:\> winget install llama-swap

# upgrade
C:\> winget upgrade llama-swap

Pre-built Binaries

Binaries are available on the release page for Linux, Mac, Windows and FreeBSD.

Building from source

1. Building requires Go and Node.js (for UI).
2. git clone https://github.com/mostlygeek/llama-swap.git
3. make clean then make to build a binary for the platform you are building on - make all builds macOS (arm64), Linux (arm64 + amd64), Windows (amd64) and FreeBSD (amd64) plus the test responder - make mac, make linux, make windows or make freebsd builds one OS, make darwin-arm64, make linux-amd64, make windows-amd64, ... builds one architecture - cross compile a single binary with make GOOS=linux GOARCH=arm64 native - Intel Mac binaries are not built here, get them from the release page
4. look in the build/ subdirectory for the llama-swap binary

Configuration

# minimum viable config.yaml

models:
  model1:
    cmd: llama-server --port ${PORT} --model /path/to/model.gguf

That's all you need to get started:

1. models - holds all model configurations
2. model1 - the ID used in API calls
3. cmd - the command to run to start the server.
4. ${PORT} - an automatically assigned port number

Almost all configuration settings are optional and can be added one step at a time:

• Advanced features
• matrix to run concurrent models with a custom swap logic DSL
• hooks to run things on startup
• macros reusable snippets
• Model customization
• ttl to automatically unload models
• unloadTimeout to tune graceful unloads (manual, API and ttl expiry)
• aliases to use familiar model names (e.g., "gpt-4o-mini")
• env to pass custom environment variables to inference servers
• cmdStop gracefully stop Docker/Podman containers
• useModelName to override model names sent to upstream servers
• ${PORT} automatic port variables for dynamic port assignment
• filters rewrite parts of requests before sending to the upstream server

See the configuration overview to get started, and the knowledge base for focused guides on the features people ask about most.

You can also just ask. The Help page (in the sidebar) is an agent that calls llama-swap's own documentation tools and answers questions about your configuration using the real text of config.example.yaml and the knowledge base, running entirely on a local model. Pick a tool-capable model on the Help page and ask away — see Writing the cmd for a model if the model answers without calling anything (older llama-server builds need --jinja added explicitly).

Those same tools are served as an MCP endpoint at /api/mcp, so any MCP client can ask about your configuration too. See Connecting an MCP client.

How does llama-swap work?

When a request is made to an OpenAI compatible endpoint, llama-swap will extract the model value and load the appropriate server configuration to serve it. If the wrong upstream server is running, it will be replaced with the correct one. This is where the "swap" part comes in. The upstream server is automatically swapped to handle the request correctly.

In the most basic configuration llama-swap handles one model at a time. For more advanced use cases, using a matrix allows multiple models to be loaded at the same time. You have complete control over how your system resources are used.

Reverse Proxy Configuration (nginx)

If you deploy llama-swap behind nginx, disable response buffering for streaming endpoints. By default, nginx buffers responses which breaks Server‑Sent Events (SSE) and streaming chat completion. (#236)

Recommended nginx configuration snippets:

# SSE for UI events and logs (also covers /api/events/logs)
location /api/events {
    proxy_pass http://your-llama-swap-backend;
    proxy_buffering off;
    proxy_cache off;
}

# Streaming chat completions (stream=true)
location /v1/chat/completions {
    proxy_pass http://your-llama-swap-backend;
    proxy_buffering off;
    proxy_cache off;
}

As a safeguard, llama-swap also sets X-Accel-Buffering: no on SSE responses. However, explicitly disabling proxy_buffering at your reverse proxy is still recommended for reliable streaming behavior.

Monitoring Logs on the CLI

# sends up to the last 10KB of logs
$ curl http://host/logs

# streams combined logs
curl -Ns http://host/logs/stream

# stream llama-swap's proxy status logs
curl -Ns http://host/logs/stream/proxy

# stream logs from upstream processes that llama-swap loads
curl -Ns http://host/logs/stream/upstream

# stream the HTTP access log, one line per request
curl -Ns http://host/logs/stream/http

# stream logs only from a specific model
curl -Ns http://host/logs/stream/{model_id}

# stream and filter logs with linux pipes
curl -Ns http://host/logs/stream | grep 'eval time'

# appending ?no-history will disable sending buffered history first
curl -Ns 'http://host/logs/stream?no-history'

Do I need to use llama.cpp's server (llama-server)?

Any OpenAI compatible server would work. llama-swap was originally designed for llama-server and it is the best supported.

For Python based inference servers like vllm or tabbyAPI it is recommended to run them via podman or docker. This provides clean environment isolation as well as responding correctly to SIGTERM signals for proper shutdown.