Overview
Mycelium is an AI agent platform. It connects agents to your databases, tools, channels, and documents through MCP — then governs every action they take.
Agents run in three modes — single, swarm, or graph — and compose into teams. Every tool call is policy-checked, every dollar is budget-capped, and every action is fully audited.
- Connect anything through MCP.
- Run single agents, swarms, and graphs — governed.
- Enterprise-grade isolation, identity, and audit.
Getting started
Create an account and a workspace for your team. A workspace keeps your sources, agents, and members separate from everyone else's.
Connect your first data source through MCP, then deploy an agent — every action will be policy-checked and budget-capped automatically.
- Sign up, then create or join a workspace.
- Connect your first MCP data source.
- Deploy your first governed agent.
Connecting your data
Agents reach your systems through MCP — the Model Context Protocol. Connect databases, channels, repositories, and tools, each pre-configured with governance-safe defaults.
Need something custom? Any MCP server — stdio or HTTP — connects in minutes. Define the tools, set the policy, deploy.
- MCP catalog: Postgres, Slack, GitHub, Brave Search, and more.
- Any custom MCP server in minutes.
- Credentials injected from vault — agents never see them.
Putting agents to work
Deploy agents in three modes: single (one autonomous agent), swarm (a self-organizing mesh), or graph (a visual workflow). Each step is governed — policy-checked, budget-capped, and traced.
Agents compose into teams: a single agent can delegate to a swarm, a graph node can trigger another graph, a swarm worker can hand off to a specialist. All on a shared durable task board.
- Single, swarm, and graph — three agent types.
- Compose into teams — any combination, any hierarchy.
- Every delegation governed, every subtask traced.
Security & governance
Every dangerous capability is intercepted. The agent holds no keys and reaches the outside world only through two mediated channels — LLM through the secret proxy, tools through the gateway.
Taint tracking blocks dangerous actions after untrusted input. Budget hard-stops prevent overspend. Human approvals gate sensitive operations. Everything is audited.
- Zero secrets exposed to agents — ever.
- Taint → sink protection against prompt injection.
- Full audit trail on every action.
Observability & oversight
See exactly what your agents are doing. Every tool call, generation, and policy decision is recorded — with timing, tokens, cost, and outcome.
Stream traces to the Trace Explorer, export to OTLP or Langfuse. Run evaluations on a schedule to catch quality drift.
- Live trace explorer with span-level detail.
- Export to OTLP or Langfuse.
- Scheduled evaluations with drift detection.