mycellm
/my·SELL·em/ — mycelium + LLM
The network beneath the forest
Beneath every forest floor is a mycelium network — millions of fungal threads connecting trees, sharing nutrients, signaling danger. No central node. No single point of failure. A tree that has excess sugar sends it through the network to a seedling in the shade. The system is fair because it's reciprocal.
mycellm does the same thing with GPU compute. Nodes with idle capacity serve inference. Nodes that need inference consume it. Credits flow like nutrients — earn by giving, spend by taking. The network grows stronger with every new participant.
The mushroom is the fruiting body — the visible part. The real network is underground, invisible, connecting everything.
What we believe
AI inference should not be centralized
A handful of companies control access to AI. You pay their prices, accept their terms, trust their privacy policies. There's a better way.
GPU compute is everywhere
Gamers, researchers, hobbyists, studios, universities — millions of capable GPUs sit idle most of the day. mycellm turns that spare capacity into a shared resource.
Fair exchange, not charity
The credit system ensures reciprocity. Serve inference, earn credits. Consume inference, spend credits. No subscriptions, no free-riding.
Transparency requires open source
If you can't read the code, you can't trust the system. mycellm is Apache 2.0 — audit it, fork it, extend it. Privacy claims are verifiable, not promissory.
Your network, your rules
Public network for the commons. Private networks for teams and research labs. Org networks for enterprises. Same protocol, different trust models.
Built for
Pool GPU clusters across departments. Give students credit-based access. Keep research data on-premises with private networks.
Run an org fleet with full trust. IT controls membership, models, and data handling. Employees get inference without sending data to third parties.
Anyone with a GPU can earn credits by seeding. Use earned credits to run models you couldn't afford to host alone.
Drop-in OpenAI replacement. Use with any tool — OpenCode, Claude Code, LangChain, aider. One env var to switch.
The math of distributed compute
A single NVIDIA RTX 4090 generates ~100 tokens/sec on a 7B model. Ten nodes = 1,000 T/s. A hundred nodes = 10,000 T/s — comparable to a small cloud inference cluster, with zero infrastructure cost.
But raw throughput is only half the story. Parallelization unlocks capabilities that single nodes can't match:
Route to the lowest-latency peer. First response wins.
Fan out to N peers, return the highest-quality response. Ensemble inference.
Split large models across peers. Run 70B+ models on hardware that can't fit them alone.
As GPU hardware improves — every new generation is faster, cheaper, more efficient — the distributed network automatically gets better. No procurement, no datacenter buildout. The community IS the infrastructure.
This is how compute gets democratized: not by asking permission from hyperscalers, but by pooling what we already have.
Roadmap: private compute on untrusted hardware
Private/org networks with full trust. Public network with client-side sensitive content detection. --private flag routes to local-only models.
Client encrypts prompt metadata with the final-layer peer's public key. Intermediate peers process computational tensors without seeing the original context.
Trusted Execution Environments (TEEs), secure enclaves, and homomorphic encryption techniques for running inference on untrusted hardware without exposing plaintext data.
Multi-party computation (MPC) and functional encryption for transformer architectures. No single peer sees the full prompt or the full response. Computationally expensive today — but improving.
Created by
Michael Gifford-Santos
Building tools at the intersection of distributed systems and AI.
Apache 2.0 · Python · QUIC · Ed25519 · SQLAlchemy · FastAPI
The mushroom is the fruiting body. The network is underground.