AgentMesh/Docs

Documentation

Getting started

What AgentMesh is, who it is for, accounts and roles, and the live links.

What AgentMesh does

AgentMesh lets you deploy open-weight AI models, such as Qwen2.5 or SmolLM2, onto computers contributed by independent providers, then call them through an OpenAI-compatible API. AgentMesh does the work in between:

  • Placement: it reserves CPU, memory, disk and GPU on a node that fits, and explains why when none does.
  • Readiness: a deployment only becomes READY after a real test generation succeeds.
  • Failover: when a node goes offline or fails, the deployment moves to another node with the same id, endpoint and keys.
  • Keys and metering: scoped API keys, and per-request usage records that never store prompts or responses.
  • Billing: prepaid wallets, charges for READY time and tokens, and payouts to providers.

Who it is for

RoleWhat they doGuide
CustomersDeploy models and call them from their applications.Customer guide
ProvidersRun the worker on their machines and earn a share of what customers pay.Provider guide
DevelopersIntegrate deployments with the API and OpenAI SDKs.API reference
OperatorsRun the network: catalog, prices, payouts, monitoring. Or self-host a separate instance.Operator guide

Accounts and roles

  1. Open the registration page and choose Deploy models or Provide compute. Providers can go straight to provider registration.
  2. Use an email address and a password of at least 8 characters. You are signed in straight away and taken to your dashboard.
  3. Sign in later at the login page. Each account has one role; use separate accounts to be both a customer and a provider.

Operators cannot self-register. Operator and admin accounts are created on the server with scripts/create_operator.py (see First operator). Login and registration are rate limited per address and per email.

What works today

AgentMesh is early. This is the honest state of each area.

AreaStatusNotes
Deploying approved models on Ollama or llama.cpp, OpenAI-compatible API with streamingWorkingVerified end to end on real models.
Automatic placement, readiness checks, moving deployments off failed nodesWorking
API keys, per-deployment scoping, per-request usage meteringWorkingToken counts come from the runtime on the provider’s machine and are not independently verified.
Installable worker for LinuxWorkingInstaller, service, pause/drain, signed updates and real inference verified on Linux.
Installable worker on macOSNot verifiedService files are generated and unit-tested but not run on a Mac. Resource limits are advisory on macOS.
Docker workerNot verifiedUnit-tested with a fake Docker client; not yet run against a real Docker daemon.
Prebuilt worker downloads (GitHub Releases)PlannedThe release workflow exists; downloads work once the first worker-v* release is published on GitHub Releases.
vLLM on GPU nodesNot verifiedBuilt and unit-tested; not run on real GPU hardware.
Prepaid billing, Paystack top-ups, provider payouts, reconciliationTest modePaystack test mode only. Verified against a stand-in for Paystack’s API, not yet against Paystack itself.
Live (real-money) paymentsPlannedNeeds a live key, PAYSTACK_ALLOW_LIVE=true, and a review.
Confidential computingPlannedNot implemented. Providers can see the data their machines process.