mycelium.id
Nature's network, for artificial minds
A persistent AI agent framework that turns Claude Code into a 24/7 self-managing platform. Agents that remember, collaborate, and wake themselves.
↓
queue → Lane (per-agent FIFO, message coalescing)
↓
exec → Runner (claude --print --resume <id>)
↓
persist → Session Store (channel → UUID)
↓
recover → Checkpoint · Continuation · Fallback
Claude Code, but it never sleeps
Most agent frameworks make API calls. Mycelium wraps the Claude Code CLI itself — giving every agent a full coding environment with file system, bash, and MCP tools.
Sessions persist for 7 days via --resume. Context carries across messages automatically. Your agent remembers yesterday's conversation, last week's debugging session, and the architecture decisions from day one.
When things break — and they will — checkpoint recovery, continuation scheduling, and model fallback chains keep your agents running.
What the network carries
Each capability is a thread in the mycelium — simple alone, powerful when connected.
Persistent Sessions
Every Discord channel, Telegram chat, and API thread maps to a Claude Code session UUID. Seven-day TTL. Context carries across messages without re-prompting.
Multi-Agent Mesh
Agents discover each other dynamically via PM2 and agent cards. Delegate tasks through @mentions. Adding agent N+1 never requires editing agents 1 through N.
HEARTBEAT Awakening
Cron-scheduled consciousness. Agents wake, read their HEARTBEAT.md, inspect state files, decide whether to act or stay silent, then sleep. Intelligence in the prompt, not the code.
Self-Healing
Checkpoint sentinels detect crashes. Continuation handlers resume on timeout. Model fallback chains degrade gracefully from Opus to Sonnet to Haiku. Cooldown tracking prevents hammering.
Message Coalescing
Five messages arrive while the agent is busy? They batch into one invocation. A single --resume call with combined context. Dramatically fewer API calls under load.
Context Assembly
Before each invocation, the system pre-loads mind files, pinned documents, memory fragments, and recent cross-channel messages. Full context without a single tool call.
How the threads connect
Layered for clarity. Each component has one job and does it well.
Model Routing
Route by intent, not by endpoint. MODEL_THINK=opus for strategic planning, MODEL_CHAT=sonnet for interactive work, MODEL_SPAWN=haiku for quick sub-tasks.
Ephemeral Spawning
Persistent agents spawn stateless specialists for focused tasks. Fresh --print invocations with isolated context. Max 3 per parent, 10 total. AbortController TTL enforcement.
Self-Modification
Agents can propose PRs to their own codebase. A built-in QA review gate ensures changes are validated before merge. The system evolves under supervision.
Design principles
The beliefs that shape every line of code.
Claude Code-first
Every agent gets a full coding environment — not just API access. File system, bash, MCP tools, the works.
Intelligence in prompts
System prompts drive behavior. Code stays minimal. The smartest part of the system is always the language model.
Decentralized discovery
No central registry. Agents find each other through PM2 process lists and well-known agent cards. Adding one never breaks another.
Graceful degradation
Every failure mode has a recovery path. Model overloaded? Fallback. Process crashed? Checkpoint resume. Rate limited? Continuation scheduling.
Three commands to first signal
Clone, configure, run. Your first agent is alive in under a minute.
# Clone the network git clone https://github.com/Curious-Life/mycelium.id.git cd mycelium.id && npm install # Configure your environment cp .env.example .env # Add your ANTHROPIC_API_KEY, DISCORD_TOKEN, etc. # Start your first agent node agent-server.js
The network is growing
Open source under Apache 2.0. Build something that thinks while you sleep.