Assistant and MCP

The optional assistant helps you explore articles already stored in RSSMonster. Open chat from the desktop toolbar or mobile Options sheet. Try requests such as “Find unread security articles from this week” or “Summarize this article.” Responses depend on available source material and the configured model; follow article links to check details against the source.

What the assistant can access

Authenticated tools can search articles, fetch selected article content, list feeds and categories, find tags, and retrieve favorite, clicked, or hot articles. Search results are bounded and scoped to your account. Content retrieval works on explicitly selected articles; it does not read your entire archive in one call. The assistant also has a crawl tool that can request a feed refresh.

The ordinary search field accepts search expressions, not arbitrary natural-language instructions. The assistant translates conversational requests into tool calls. It cannot retrieve articles that RSSMonster has never collected just by searching your archive.

Enable the assistant

Configure model providers using Model Usage, then enable chat in server/.env:

INFERENCE_AI_ENABLED=true
INFERENCE_ASSISTANT_ENABLED=true
INFERENCE_AGENT_TIMEOUT_MS=300000
# Optional; use a reasoning effort supported by the selected model.
ASSISTANT_REASONING_EFFORT=

The assistant uses the openai-compatible adapter for OpenAI or OpenAI-compatible gateways, including local Ollama. Put its credentials and model settings in inference/.env:

ASSISTANT_API_KEY=your-openai-api-key
ASSISTANT_BASE_URL=https://api.openai.com/v1
ASSISTANT_PROVIDER=openai-compatible
ASSISTANT_MODEL=gpt-4o-mini
# Optional gateway endpoint; choose a model available at that endpoint.
# ASSISTANT_BASE_URL=http://127.0.0.1:11434/v1

Model calls run through RSSMonster’s configured inference provider using Chat Completions. A gateway does not need Responses API support, but the assistant model must support tool calling and streaming. For local Ollama, use a non-empty placeholder key such as ollama and an installed model name. See the inference gateway guide for setup and Compose forwarding requirements. The optional reasoning effort setting belongs in server/.env; leave it blank unless needed by your model.

Local embeddings, classification, summaries, scoring, Smart Folder recommendations, and feed rediscovery do not require this key when configured with Qwen and ModernBERT. Chat is an independent optional capability; enabling it does not schedule background summaries, inferred tags, or article scoring.

For the comprehensive MySQL Compose profile, put the assistant values and INFERENCE_ASSISTANT_ENABLED=true in the repository-root .env; that profile already forwards each value to the responsible service. Recreate it with docker compose -f docker-compose.mysql.yml up -d. For source installations, restart inference and the server. The client learns chat availability from the server, so these server-side flags do not require a client rebuild.

Provider credentials belong in inference configuration. RSSMonster executes user-scoped tools locally; inference performs model calls. When using an external provider, prompts and the source material included in those requests leave the RSSMonster host. Embedding and assistant providers can have different settings.

If chat is unavailable, check both enable flags, inference readiness, and provider configuration. Timeouts and capability circuit breakers are described in Configuration.

Integrate using MCP

The Model Context Protocol transport is mounted at /mcp, outside /api. It accepts authenticated GET and POST requests using a JWT bearer token and exposes the shared RSSMonster tools. Obtain a JWT through /api/auth/login and send it as Authorization: Bearer <token>. /api/agent is the built-in assistant endpoint; there is no /api/mcp route. MCP authentication and tools do not themselves require enabling the built-in assistant’s model provider.

A compatible integration must support the transport and bearer authentication; ordinary RSS clients should use Fever or Google Reader. See RSSMonster API for login and rate limits. The implementation in server/controllers/mcp.js is the reference for tool names and schemas.