Note: For security reasons, features that require reading or writing external files must be **disabled by default**. This covers features like: MCP, model save/load
Out-of-scope features:
- Backend:
- Features that require a loop of external API calls, e.g. server-side agentic loop. This is because external API calls in C++ are costly to maintain. Any complex third-party logic should be implemented outside of server code.
- Features that expose the internal state of the model to the API, example: getting the intermediate activation from API. This is because llama.cpp doesn't support a stable API for doing this, and relying on `eval_callback` can make it complicated to maintain as this API is not intended to be used in multi-sequence setup.
- Model-specific features. All API calls and features must remain model-agnostic.
- Frontend:
- Third-party plugins, it is costly to maintain a public plugin API for such features. Instead, users can make their own MCP server for their needs.
- Customizable themes, it is also costly to maintain. While we do focus on the aesthetic, we try to achieve this by perfecting a small set of themes.
- Browser-specific features, example: [Chrome's built-in AI API](https://developer.chrome.com/docs/ai/built-in-apis).
- **Inference mode**: The default mode for performing inference with a single loaded GGUF model.
- **Router mode**: Enables management of multiple inference server instances behind a single API endpoint. Requests are automatically routed to the appropriate backend instance based on the requested model.
The core architecture consists of the following components:
-`server_context`: Holds the primary inference state, including the main `llama_context` and all active slots.
-`server_slot`: An abstraction over a single “sequence” in llama.cpp, responsible for managing individual parallel inference requests.
-`server_routes`: Middleware layer between `server_context` and the HTTP interface; handles JSON parsing/formatting and request routing logic.
-`server_http_context`: Implements the HTTP server using `cpp-httplib`.
-`server_queue`: Thread-safe queue used by HTTP workers to submit new tasks to `server_context`.
-`server_response`: Thread-safe queue used by `server_context` to return results to HTTP workers.
-`server_response_reader`: Higher-level wrapper around the two queues above for cleaner code.
-`server_task`: Unit of work pushed into `server_queue`.
-`server_task_result`: Unit of result pushed into `server_response`.
-`server_tokens`: Unified representation of token sequences (supports both text and multimodal tokens); used by `server_task` and `server_slot`.
-`server_prompt_checkpoint`: For recurrent (e.g., RWKV) and SWA models, stores snapshots of KV cache state. Enables reuse when subsequent requests share the same prompt prefix, saving redundant computation.
-`server_models`: Standalone component for managing multiple backend instances (used in router mode). It is completely independent of `server_context`.
The server context maintains a single batch shared across all slots. When `update_slots()` is invoked, the system iterates through all active slots to populate this batch. For each slot, either a generated token from the previous decoding step or available prompt tokens are added to the batch.
Batching constraints apply: slots can only be batched together if they share compatible configurations. For instance, slots using a specific LoRA adapter can be batched with each other, but not with slots using a different LoRA adapter or no adapter at all.
Once the batch reaches capacity or all slots have been processed, `llama_decode` is called to execute the inference. This operation represents the primary computational bottleneck in `update_slots()`.
Following decoding, the system either retrieves embeddings or samples the next token using `common_sampler_sample`. If a slot has remaining prompt tokens to process, it yields until the next `update_slots()` iteration.
`server_context` runs on a dedicated single thread. Because it is single-threaded, heavy post-processing (especially after token generation) should be avoided, as it directly impacts multi-sequence throughput.
Each incoming HTTP request is handled by its own thread managed by the HTTP library. The following operations are performed in HTTP worker threads:
- JSON request parsing
- Chat template application
- Tokenization
- Conversion of `server_task_result` into final JSON response
- Error formatting into JSON
- Tracking of partial/incremental responses (e.g., streaming tool calls or reasoning steps)
**Best practices to follow:**
- All JSON formatting and chat template logic must stay in the HTTP layer.
- Avoid passing raw JSON between the HTTP layer and `server_slot`. Instead, parse everything into native C++ types as early as possible.
Here is an example trace of an API request for text completion:
- A request arrives at the HTTP layer.
- The request is routed to the corresponding handler inside `server_routes`. In this case, `handle_completions_impl` is invoked.
- The handler parses the input request, constructs a new `server_task`, and passes it to `server_res_generator`.
-`server_res_generator` creates a new `task_result_state` for each task:
-`task_result_state` stays in the HTTP layer, responsible for keeping track of the current state of the response (e.g., parsing tool calls or thinking messages).
-`server_task` is moved into `server_queue` inside `server_context`.
-`server_context` launches the task by moving it into an available slot (see `launch_slot_with_task()`).
-`update_slot()` processes the task as described in the "Batching" section above.
- Results may be sent using `send_partial_response` or `send_final_response`, which creates a new `server_task_result` and pushes it to the response queue.
- At the same time, `server_res_generator` listens to the response queue and retrieves this response.
- As the response is stateless, `server_res_generator` calls `response->update()` to update the response with the current state.
-`server_res_generator` then calls `response->to_json()` and passes the response to the HTTP layer.
This endpoint is intended to be used internally by the Web UI and subject to change or to be removed in the future.
**GET /tools**
Get a list of tools, each tool has these fields:
-`tool` (string): the ID name of the tool, to be used in POST call. Example: `read_file`
-`display_name` (string): the name to be displayed on UI. Example: `Read file`
-`type` (string): always be `"builtin"` for now
-`permissions` (object): a mapping string --> boolean that indicates the permission required by this tool. This is useful for the UI to ask the user before calling the tool. For now, the only permission supported is `"write"`
-`definition` (object): the OAI-compat definition of this tool
**POST /tools**
Invoke a tool call, request body is a JSON object with:
-`tool` (string): the name of the tool
-`params` (object): a mapping from argument name (string) to argument value
Returns JSON object. There are two response formats:
Format 1: Plain text. The text will be placed into a field called `plain_text_response`, example:
```json
{
"plain_text_response":"this is a text response"
}
```
The client should extract this value and place it inside message content (note: content is no longer a JSON), example
```json
{
"role":"tool",
"content":"this is a text response"
}
```
Format 2: Normal JSON response, example:
```json
{
"error":"cannot open this file"
}
```
That requires `JSON.stringify` when formatted to message content:
The project includes a web-based user interface for interacting with `llama-server`. It supports both single-model (`MODEL` mode) and multi-model (`ROUTER` mode) operation.
The SvelteKit-based Web UI is introduced in this PR: https://github.com/ggml-org/llama.cpp/pull/14839
### Features
-**Chat interface** with streaming responses
-**Multi-model support** (ROUTER mode) - switch between models, auto-load on selection
**Note:** The Vite dev server automatically proxies API requests to `http://localhost:8080`. Make sure `llama-server` is running on that port during development.
