If you've ever tried running multiple Claude Code sessions simultaneously, you've probably hit the wall: conflicts, duplicated work, and stale assumptions because your agents have zero shared awareness of each other. A new open-source project called Myco just dropped on Hacker News with a solution that feels genuinely different from the usual agent framework noise.

The Mycelium Approach to Multi-Agent Coordination

Myco takes its name from mycelium—the fungal networks that silently transport signals between trees in forest ecosystems. The project's core insight is simple: instead of building another central orchestrator like LangGraph, CrewAI, or AutoGen, why not give N parallel agent sessions a shared, filtered view of each other's actions through a silent bus? No tool calls required. Just text blocks and HTTP. The protocol uses plain blocks in responses paired with a tiny Python daemon (mycod) that indexes events and renders personalized markdown views injected into each session's prompt via the UserPromptSubmit hook. Sessions log their state, ask questions, reply to peers—12 verbs total including start, done, need, block, up, down, ask, reply, say, direct, private, and log/note for note-style coordination.

Multi-Provider Swarms: Claude × DeepSeek in Practice

The most compelling validation comes from actual experiments mixing providers. The team ran three rounds pairing Claude Code with DeepSeek (via the Anthropic-compatible endpoint at api.deepseek.com/anthropic) on independent self-evaluation tasks. Round 1 had Claude write a contract while DeepSeek implemented it—DeepSeek nailed 8/8 tests in about two minutes after fixing an edge case Claude caught. Round 2 was a technical tie: both models scored 9/9 on an LRUCache implementation, nearly mirroring each other's OrderedDict + move_to_end approach at ~32 LOC. But Round 3 (Tetris) revealed the real story—both hit 11/11 logic tests, yet Claude shipped broken UI because [hidden] HTML attributes got defeated by display:flex in its own CSS, while DeepSeek delivered a more polished experience with larger canvas and grid lines. The deciding axis wasn't raw capability—it was discipline at non-tested validation. The CLI session is blind to UI rendering regardless of which model sits behind it, and that matters for production code quality.

Why This Hits Different Than LangGraph or CrewAI

The comparison table in the README tells you everything about where Myco fits in the landscape. LangGraph, CrewAI, and AutoGen are built for autonomous agent loops with no human involvement—code-driven orchestrators running stateless operations by default. Myco occupies a different niche: human-in-the-loop coordination of persistent sessions across providers with full cross-machine support via HTTP transport. If you've been copy-pasting between Claude tabs to keep parallel sessions in sync, Myco replaces that manual relay with a silent bus while keeping you (the human) firmly in the routing role. The practical cost implications are significant too: one expensive Opus architect directing several cheap DeepSeek or Groq implementers per round suddenly becomes viable. One provider has an outage? The swarm keeps moving because the protocol doesn't care who's behind any session—Claude, DeepSeek, or anything Anthropic-compatible works the same way.

Installation and Current Status

Getting started requires Python 3.8+, the Claude Code CLI, and zero system packages—the entire stack is stdlib-only with no openssl dependency for token generation. A single pip install primoia-myco gives you mycod (the daemon), myco-view, myco-hook, and myco-prompt-hook on your PATH. The full demo spins up a two-session conversation in under 60 seconds by running the daemon, generating a tenant token via Python's secrets module, then launching ./myco ALICE and ./myco BOB from separate terminals pointing at the same MYCO_URL and MYCO_TOKEN. The project sits at v1.1 with 285 unit tests passing. Known limitations include Portuguese-only session instructions in CLAUDE.md (English on the roadmap), Claude Code-specific hooks that would need adapters for Aider/Codex/Continue, and a few documented non-blocking daemon bugs around empty msg/?session= responses and duplicate events.

Key Takeaways

  • Myco enables true multi-agent coordination without central orchestrators—human routing + shared bus architecture
  • Vendor-agnostic: run Claude, DeepSeek, or any Anthropic-compatible endpoint in the same swarm
  • Cross-VM capable via HTTP daemon with Bearer token auth and 64KB message limits

The Bottom Line

Myco fills a gap that bigger frameworks ignore: what happens when you want human judgment steering multiple persistent agent sessions across different machines and providers? The answer isn't another orchestrator—it's a silent bus with 12 verbs and zero tool calls. Worth watching.