AgentMesh
Get started
Open beta: community compute network

Run open models on compute you don't manage.

Deploy open-weight models onto a network of independent machines and call them through an OpenAI-compatible API. Placement, health checks, failover and billing are handled for you.

Start deploying Contribute compute
  • OpenAI-compatible
  • Automatic failover
  • Pay only while serving
from openai import OpenAI

client = OpenAI(
    base_url="https://agentmesh-riuz.onrender.com/api/v1/deployments/<id>/v1",
    api_key=os.environ["AGENTMESH_API_KEY"],
)

reply = client.chat.completions.create(
    model="qwen2.5:0.5b",
    messages=[{"role": "user", "content": "Summarise this ticket"}],
    stream=True,
)

qwen2.5-0.5b

Ready

  1. Scheduled

    Capacity reserved on a matching node

  2. Model verified

    Weights pinned and checked by digest

  3. Ready

    Test generation passed, endpoint live

—
Nodes online
—
GPU nodes online
—
Deployable models
—
Deployments serving

Loading live network figures…

Product

Infrastructure you don't have to run.

Everything between “I want this model” and a production endpoint: scheduling, health, failover, keys and billing.

Drop-in OpenAI API

Every deployment gets a /v1 endpoint with chat, completions and streaming. Point your existing OpenAI client at it and keep your code.

Automatic placement

The scheduler reserves CPU, memory and GPU on a node that fits, and tells you exactly why when none does.

Self-healing

When a node goes offline or fails health checks, the deployment moves to another one. Same id, same endpoint, same keys.

Verified models

Weights are pinned to exact versions and checked by digest. A deployment only goes live after a real test generation passes.

Prepaid, per ready minute

Top up with Paystack and pay for reserved resources only while your model is serving, plus tokens. Every charge is itemised.

Metering without content

Usage is recorded per request with no prompts or responses stored. API keys can be scoped to a single deployment.

How it works

A two-sided compute network.

For developers

From catalog to endpoint in minutes

  1. 1

    Pick a model

    Browse approved open-weight models with their runtimes, memory needs and hourly price.

  2. 2

    Deploy

    AgentMesh reserves capacity on a node, downloads and verifies the model, and runs a test generation.

  3. 3

    Call it

    Create a scoped API key and use any OpenAI client. Watch requests, tokens and charges live.

Deploy a model

For providers

Turn idle machines into income

  1. 1

    Install the worker

    A per-user agent or a Docker worker. Outbound connections only; no ports to open.

  2. 2

    Set your limits

    Cap CPU, memory and disk, set availability hours, and pause or drain any time.

  3. 3

    Get paid

    Earn your share of every charge for the deployments you serve, paid out to your bank via Paystack.

Become a provider

Pricing

Pay for what serves.

Prepaid and transparent. You're charged for the resources reserved for your deployment only while it is ready, plus the tokens it processes. Downloads, startup and unhealthy time are free.

Providers receive most of every charge.

No prices published yet

The network operator hasn't set a price list, so usage is currently measured but not charged.

Start deploying

Status

What works today

AgentMesh is early. Here is the honest state of the product.

Working

  • Deploy approved open-weight models (Ollama, llama.cpp) and call them through an OpenAI-compatible API, with streaming
  • Automatic placement on a node with enough free capacity, and automatic moves when a node fails
  • API keys, per-key deployment scoping, and per-request usage metering (no prompt or response content stored)
  • A Docker worker and an installable agent for Linux and macOS with resource limits, pause, drain and availability schedules
  • Provider, customer and operator dashboards backed by live data, with an audit log
  • Prepaid billing: versioned prices, charges for READY time and tokens, Paystack top-ups and provider payouts (Paystack test mode)

Planned, not available yet

  • Live (real-money) Paystack payments
  • vLLM on GPU nodes (built, not yet verified on real GPU hardware)
  • Confidential computing (hardware-isolated nodes with attestation)

Trust model

Models run on computers owned by independent providers. A provider can inspect the model files, the process memory, and every prompt and response its node handles. Container isolation protects the provider's machine; it does not hide your data from the provider. Don't send confidential data to community nodes.

Ship your first model today.

Start with a small model like SmolLM2 or Qwen2.5 and have an OpenAI-compatible endpoint in minutes.

Get started Become a provider