Self-hosted AI agents for Mattermost that act on your behalf through tools and MCP servers, plus a built-in Mattermost MCP server for external AI clients such as Claude Code, Claude Desktop, Cursor, and VS Code.
Mattermost Agents (plugin ID mattermost-ai) adds AI agents to a Mattermost server. You talk to an agent in a direct message, the Agents pane, or a channel. The agent can search and post in Mattermost, call tools on other systems through the Model Context Protocol (MCP), and work with files, all with the permissions of the person it is acting for. You choose the LLM, including models that run on your own infrastructure, so with a self-hosted model, conversations and data stay in your environment.
Does Mattermost have an official MCP server? Yes. It is built into this plugin and needs no separate installation. After an admin enables it, MCP clients connect to /plugins/mattermost-ai/mcp-server/mcp on your Mattermost server and sign in as a Mattermost user. See Mattermost MCP Server.
- Agents that act on your behalf: Chat with a personal agent that searches and reads Mattermost, posts and sends messages, schedules posts and reminders, and uses tools from other systems. Configure multiple agents with their own instructions, models, and tool access. See Use Mattermost Agents as your personal AI assistant.
- Tools and MCP: Agents use built-in Mattermost tools, remote MCP servers with per-user OAuth (with a Connect card in the conversation), and tools registered by other Mattermost plugins. Dynamic tool loading keeps large tool catalogs out of the model's context.
- Built-in Mattermost MCP server: Connect Claude Code, Claude Desktop, Cursor, VS Code, or any MCP client that supports streamable HTTP to Mattermost, using OAuth with dynamic client registration or a personal access token. Over 100 tools run with the signed-in user's own permissions. See Mattermost MCP Server.
- Governance: Per-tool approval policies, user, team, and channel access controls for agents, attribute-based access control, server audit logging, token usage tracking, and OpenTelemetry tracing. Admins decide which tools exist and which need approval. See the Admin Guide.
- Bring your own LLM: Use local models through OpenAI-compatible APIs (Ollama, vLLM), or OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google Gemini and Vertex AI, Cohere (including Cohere North), and Mistral. See Provider configuration and Sovereign AI deployments.
- Summaries, search, and meetings: Summarize threads, unread messages, and channels; ask questions about a channel; search with keyword and semantic search; and summarize Mattermost Calls and Zoom meeting recordings.
- Files and artifacts: Agents create text files attached to their replies, read attachments, analyze images with vision-capable models, and render HTML artifacts in the conversation.
- Channel use: @mention an agent in a channel, or set a channel to auto-reply. Tool results in channels can be shared or kept private by the person who asked.
- Custom prompts: Save reusable prompts, pin them in the Agents pane, and share them with your team.
Some features depend on your Mattermost license level. See the license requirements.
Documentation is in the docs directory. The same pages are published on docs.mattermost.com.
For users
- Use Mattermost Agents: Start a conversation, choose an agent, attach files, and find every user feature
- Use Agents as your personal AI assistant: What an agent can do across channels, MCP servers, and integrations, and how to prompt it
- Use tools and approve actions: Accept or reject tool calls and connect your accounts
- Use agents in channels: Mentions, tool use in channels, and auto-reply
- Summaries and AI search: Summarize threads and channels, and search with AI
- Meeting summaries: Summarize Calls and Zoom meetings
- Custom prompts: Save, share, and pin reusable prompts
- What agents can see and keep: What an agent reads, what it can reach, and what is stored
For admins
- Admin guide: Requirements, quick start, settings map, and license requirements
- Configure LLM services, Provider configuration, and AWS Bedrock setup
- Create and manage agents and Control who can use agents
- Connect agents to MCP servers and Tool approval and credentials
- Control agents in channels
- Web search and Set up semantic search
- Secure AI-generated links and HTML
- Monitor Agents, Back up and automate, and Troubleshooting
- Data handling and Sovereign AI deployments
- Upgrade from Agents 1.x
For developers
- Mattermost MCP server: Connect Claude Code, Claude Desktop, Cursor, VS Code, and other MCP clients to Mattermost
- Mattermost MCP tool reference: All 121 built-in tools
- LLM Bridge for plugins: Call agents and services from another plugin
- Load testing: Configure the in-process mock LLM and the load-test-ng Agents controller
- Download the latest release from the releases page. You can also download the experimental latest master
- Upload and enable the plugin through the Mattermost System Console
- Configure your desired LLM provider settings
- Mattermost Server v12.0.0+
- PostgreSQL database, with the pgvector extension for semantic search
- Site URL set on the Mattermost server
- Network access to your chosen LLM provider
After installation, complete these steps to get started:
- Go to System Console > Plugins > Agents and select Add an AI Service. Choose your LLM provider and enter its credentials and a default model.
- Open the Agents page from the product switcher (grid icon) and select Create agent. Choose a username and the service.
- Optionally set who can use the agent on its Access tab.
- Send the agent a direct message, or open the Agents pane from the apps bar.
For more, see the Admin guide.
The plugin provides a Go client library for other Mattermost plugins and the Mattermost server to interact with the AI plugin's LLM Bridge API. This allows you to easily add AI capabilities to your own plugins or server-side features.
See the Bridge Client README for installation and usage instructions.
- Go 1.27+
- Node.js 24.11+
- Access to an LLM provider (OpenAI, Anthropic, etc.)
-
Setup your Mattermost development environment by following the Mattermost developer setup guide. If you have a remote mattermost server you want to develop to you can skip this step.
-
Setup your Mattermost plugin development environment by following the Plugin Developer setup guide.
-
Clone the repository:
git clone https://github.com/mattermost/mattermost-plugin-agents.git
cd mattermost-plugin-agents- Optional. If you are developing to a remote server, setup environment variables to deploy:
MM_SERVICESETTINGS_SITEURL=http://localhost:8065
MM_ADMIN_USERNAME=<YOUR_USERNAME>
MM_ADMIN_PASSWORD=<YOUR_PASSWORD>- Run deploy to build the plugin
make deploy- Run
make helpfor a list of all make commands - Run
make check-styleto verify code style - Run
make testto run the test suite - Run
make e2eto run the e2e tests - Run
make evalsto run prompt evaluations interactively (with TUI) - Run
make evals-cito run prompt evaluations in CI mode (non-interactive)
The streaming code has benchmark tests to measure performance and detect regressions:
# Run all streaming benchmarks
go test -bench=. -benchmem ./llm/... ./streaming/...
# Run specific benchmark
go test -bench=BenchmarkStreamToPost -benchmem ./streaming/...
# Run with CPU profiling
go test -bench=BenchmarkReadAll -cpuprofile=cpu.prof ./llm/...Benchmarks cover:
ReadAll()stream consumption with varying response sizesTokenUsageLoggingWrapperinterception overheadStreamToPost()full processing pipeline with WebSocket events
The evaluation system supports testing with multiple LLM providers: OpenAI, Anthropic, and Azure OpenAI. By default, evaluations run against all providers, but you can target specific ones:
# Run with all providers (default)
make evals
# Run with only Anthropic
LLM_PROVIDER=anthropic make evals
# Run with OpenAI and Azure
LLM_PROVIDER=openai,azure make evals
# Use a specific model
ANTHROPIC_MODEL=claude-opus-5-5 make evalsSee cmd/evalviewer/README.md for complete documentation on environment variables and configuration options.
This repository is licensed under Apache-2, except for the enterprise directory which is licensed under the Mattermost Source Available License. See Mattermost Source Available License to learn more.
