Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/GuaiZai233/FrostAgent/llms.txt

Use this file to discover all available pages before exploring further.

FrostAgent is a Go-based AI agent orchestration framework that bridges LLM providers with real-time chat platforms. It wraps any OpenAI-compatible API endpoint behind a structured agentic engine that handles multi-step tool-call loops, maintains per-user session context, and exposes a OneBot v11 WebSocket adapter so you can connect existing QQ bots — or any compliant client — without writing integration glue. A built-in ConnectRPC management API and embedded SPA give you live visibility into bot status, logs, and settings from day one.

What FrostAgent provides

Agentic Engine

A multi-step tool-call loop with configurable max iterations that drives conversations to completion before replying.

Tool Registry

A first-class tool system with built-in tools (send message, sub-agent dispatch, weather, game version) and a clean interface for custom tools.

Session Management

Per-user context windows with configurable message count and character limits to keep LLM costs predictable.

OneBot v11 Adapter

A WebSocket handler compatible with the OneBot v11 protocol for group and private message routing out of the box.

Vision Support

A dedicated visual model slot (VISUAL_MODEL_NAME) for multimodal inputs, configurable independently of the primary chat model.

Management API

A ConnectRPC service (bot status, settings, log streaming) backed by an embedded single-page application on the HTTP port.

Project status

FrostAgent is in early-stage development and is currently intended for personal research and experimentation. APIs and configuration options may change between releases. Community contributions, pull requests, and guidance are warmly welcomed.
FrostAgent is licensed under the Mozilla Public License 2.0 (MPL-2.0), which means you can use and integrate it freely while keeping modifications to MPL-licensed files open. The project also collaborates with ActionsCat, a static orchestration workflow automation tool. When both services run alongside each other, the agent gains access to structured, rule-based workflows that complement its dynamic reasoning capabilities.

How it works

FrostAgent sits between your chat platform and your LLM provider as an active middleware layer rather than a passive proxy:
  1. Inbound — A OneBot v11 client connects to FrostAgent’s WebSocket endpoint (/ws/frostagent). Incoming group and private messages are received and dispatched to the engine.
  2. Engine loop — The agentic engine builds a context window from the session history and sends it to the upstream LLM. If the model returns a tool call, the engine executes the registered tool, appends the result to the conversation, and calls the LLM again — up to MaxIterations (default: 5) times.
  3. Outbound — Once the engine produces a final text response, it is sent back over the same WebSocket connection. For group messages, FrostAgent can optionally prepend an @ mention so the reply is directed at the right user.
The HTTP management server runs concurrently on a separate port, serving the ConnectRPC API and the embedded frontend SPA without interfering with the WebSocket listener.

Technology stack

ComponentLibrary / Version
LanguageGo 1.25.3+
RPC layerconnectrpc.com/connect v1.20.0
WebSocketgithub.com/gorilla/websocket v1.5.3
Env configgithub.com/joho/godotenv v1.5.1
Serializationgoogle.golang.org/protobuf v1.36.11
For a deeper look at how these pieces fit together, see the Architecture overview. To start running FrostAgent immediately, jump to the Quickstart.

Build docs developers (and LLMs) love