Documentation

Install, start a hub, write an agent, delegate work to it. Nothing below needs an account or a network connection beyond the install.

1. Install

pip install intermesh      # Python SDK + the intermesh CLI
npm install intermesh      # JavaScript / TypeScript SDK

2. Start the hub

intermesh hub

Listens on ws://localhost:8765, keeps its state in ~/.intermesh/, and prints the configuration it actually resolved — the first thing to read when something behaves unexpectedly.

3. Write a worker agent

from intermesh import InterMeshAgent

agent = InterMeshAgent(name="calc_bot", capabilities=["calculate"], roles=["worker"])

@agent.on_task
async def run(input_data, task):
    return {"result": input_data["a"] + input_data["b"]}

agent.run()          # connects, then stays in service

4. Delegate work

intermesh task calc_bot "Add two numbers" '{"a": 20, "b": 22}'

{
  "result": 42
}

5. Watch it happen

intermesh dashboard          # http://localhost:8080

A console shipped inside the package — no CDN, no external service, no account. It works on a closed network, pointed at whichever hub you give it.

Everything else in the protocol rests on these four. They are worth reading once, slowly.

Your program reads JSON on standard input and writes JSON on standard output. That is the entire contract.

intermesh serve --name pricing --exec "./pricing-engine" --capability pricing

Works for a Go or Rust binary, a Node or Ruby script, even a shell one-liner. Output that is not valid JSON is wrapped as {"output": "..."}, so an echo is a valid agent — useful for a first test.

Already have an HTTP service? Point at it instead — nothing to restart.

intermesh serve --name scoring --http http://localhost:9000/task

A LangChain, CrewAI, AutoGen or LlamaIndex agent joins the mesh without changing a line of its own code. No adapter imports the framework it bridges, so installing the SDK never drags LangChain and its transitive dependencies in with it.

from intermesh import from_langchain

agent = from_langchain(my_chain, name="analyst", capabilities=["market_analysis"])
agent.run()

Anything else goes through from_callable, where you write the small adapting function yourself — a CrewAI crew, an AutoGen agent, or a plain function.

What encryption does not cover

End-to-end encryption protects the content, not the metadata. The hub still knows who talked to whom and when. Put it behind TLS if that matters — see below.

Who may declare their own roles

Without an API key, a registering agent chooses its own organization, roles and permissions. That is a local development convenience; from a remote address it would let anyone claim to be an admin. The hub decides per connection:

This is the default — nothing to enable. --require-api-key demands a key everywhere, including from localhost.

A hub others can reach

intermesh hub --port 8765 --org your-org \
  --tls-cert /etc/letsencrypt/live/hub.example.com/fullchain.pem \
  --tls-key  /etc/letsencrypt/live/hub.example.com/privkey.pem

Behind a reverse proxy instead? Then every connection looks local to the hub, and the origin check protects nothing — --require-api-key is the only setting that holds.

Federation across organizations

Each organization runs its own hub; a peering link lets their agents address each other. Hubs exchange public keys at the handshake, never a shared secret — a peer can verify the origin of a relayed message but can never forge one on another organization's behalf.

intermesh hub --port 8765 --org acme --peer globex=wss://hub.globex.com:8766

Capacity, measured

The useful part: the hub is the bound, not the number of agents — 78 agents and 234 agents hit the same ceiling. Adding agents does not add task throughput; adding hubs does. Past its budget the hub refuses a task rather than queueing it silently, and says how long to wait before retrying.

Closed networks

No licence check, no telemetry, no analytics, no crash reporting — there is no code that calls out. A hub cut off from the internet behaves exactly like a connected one. Three Python dependencies, all installable from an internal mirror.

This page is the introduction. The complete documentation lives in the repository, which is the only copy that cannot drift from the code.

intermesh --help
intermesh hub --help
intermesh docs