Networks: public, private, federated

mycellm is shaped like the internet, not like a service.

There is one public network anyone can join, any number of private networks — your own devices, an affinity group, a company or an org — and federation between them, with each side deciding what it exposes.

        PUBLIC NETWORK                 PRIVATE NETWORKS
   ┌─────────────────────┐        ┌──────────┐  ┌──────────────┐
   │  anyone may join    │◀─────▶ │ your     │  │ your org     │
   │  open membership    │  fed   │ devices  │  │ or team      │
   │  rate-limited       │        └──────────┘  └──────────────┘
   └─────────────────────┘              ▲              ▲
                                        └──── fed ─────┘

That shape is the product. Not “cheaper tokens”, and not “inference on hardware you own” — who constitutes the network, and who governs what crosses its boundary.

Why this is the distinguishing property

Every alternative is one network with one operator, and you are a customer of it. That is true of the centralised APIs, and it is equally true of the decentralised marketplaces — a marketplace is still a single network whose membership and policy someone else sets. Encryption changes who can read your prompt; it does not change whose network it crossed or who decided the rules.

mycellm lets you stand up a network and interconnect it with others on terms you set:

• Public — join the open network, serve or consume, earn credits.
• Private — your own devices, a group of friends, a company. Own bootstrap, own membership, own credit ledger, own trust policy.
• Federated — link networks with per-network model scope, so a model visible to your team need not be visible to the public network you also belong to.

A node can belong to several at once, and each membership carries its own policy. That is why the interesting questions in mycellm are constitutional rather than commercial.

Governance is enforced, not promised

Each of these is implemented, and refuses visibly rather than silently:

Control	What it does
Egress policy per target	Whether a prompt may leave this machine at all, decided per destination — not a global switch
Trust ceilings	local / trusted / any — a request declares how far it may travel, and the planner honours it
Per-network model scope	A model advertised on one network stays invisible on another
Serving groups	Which models a node offers, and to whom
Visible refusals	Every blocked target appears in the execution plan with its reason. A target blocked by policy is otherwise indistinguishable from one that was never there

Ask a node what it would do, without running anything:

curl -X POST localhost:8420/v1/node/plan \
  -H "Authorization: Bearer $MYCELLM_API_KEY" \
  -d '{"model": "", "messages": [{"role": "user", "content": "hello"}]}'

The response includes rejected — the targets considered and refused, each with a cause. That list is the point: a fabric that silently picks a route is unauditable.

Where this sits among the alternatives

Worth stating plainly, including where others are ahead.

Darkbloom (Eigen Labs) routes encrypted inference to idle Apple Silicon Macs, around 50% below centralised pricing, and is live on OpenRouter. It is well funded, further along on volume, and genuinely private in the cryptographic sense — Secure Enclave-rooted keys and per-response attestation. If what you want is cheap private inference from a pool, it is a serious option, and mycellm is not trying to beat it on price.

The difference is structural rather than a feature comparison:

	A marketplace	mycellm
Networks	One, operated by the vendor	Public + any number of private, federated
Membership	Set by the operator	Set by whoever stands up the network
Your hardware runs	Other people’s prompts	Whatever your network’s policy says
Devices	Mac-only, ≥24GB (Darkbloom)	Macs, Linux, iPhone and iPad
Privacy model	Cryptographic — proving a stranger’s machine behaved	Constitutional — you set the boundary, and crossings are logged and refusable

Ollama / LM Studio own the single-box experience and deserve to; mycellm should not be your choice for one machine. exo is the closest technical sibling for Mac clusters. OpenRouter is a distribution channel, not a competitor.

What this does not claim

• Not cheaper than centralised APIs. It is not trying to be.
• Not faster. A fleet of personal devices trades latency for reach; the job queue is how that trade is made honest instead of hidden.
• Not better answers from more models. The swarm strategy is available and measured, and on bounded tasks it has not yet beaten a single good model — which is why nothing here claims that it does.