cosmonapse

Event-driven Agent-to-Agent protocol SDK for Python.

Python 3.11+ · Apache-2.0

Install

pip install cosmonapse

# Provider-backed Neurons (Ollama / HuggingFace / OpenAI / Anthropic / groq /
# openrouter / together / mistral) need httpx
pip install httpx

This puts both import cosmonapse and the cosmo CLI on your PATH. The CLI has a single implementation - this package - and is also reachable from npm: npm install -g @cosmonapse/sdk ships a launcher that delegates to python -m cosmo, auto-installing this package into a private environment on first run if needed. One CLI build either way.

Quick start

The fastest path is init -> scaffold -> code:

cosmo init my-app -n demo
cd my-app
python demo.py     # one process, in-process bus - no setup

cosmo init writes the standard skeleton every example follows - config.py, neurons/, effector/, brain.py, demo.py. Then you code: new Neuron modules under neurons/, tool families under effector/, wiring in brain.py. Or wire everything by hand:

import asyncio
from cosmonapse import (
    Axon, Dendrite, Neuron,
    MemoryRegistryStore,
    connect_synapse,
)

async def main():
    # 1. Connect to a running synapse (cosmo synapse start memory)
    synapse = await connect_synapse("cosmo://127.0.0.1:7070")
    try:
        # 2. Define a pure agent function (zero protocol knowledge)
        async def my_agent(input, context):
            return {"answer": f"You asked: {input['q']}"}

        # -- Worker Dendrite: hosts an Axon --------------------------
        worker = Dendrite(synapse=synapse, namespace="demo")
        worker.attach_axon(Axon(neuron_id="answerer", neuron_fn=my_agent))

        # -- Orchestrator Dendrite: drives the workflow ---------------
        orch = Dendrite(
            synapse=synapse,
            registry_store=MemoryRegistryStore(),
            namespace="demo",
        )

        @orch.on_agent_output
        async def done(sig):
            print("Got answer:", sig.payload["output"])
            await orch.emit_final(
                trace_id=sig.trace_id,
                parent_id=sig.id,
                result=sig.payload["output"],
            )

        async with orch, worker:
            await orch.dispatch_task(neuron="answerer", input={"q": "hi"})
            await asyncio.sleep(0.5)
    finally:
        await synapse.close()

asyncio.run(main())

Start the dev synapse first:

cosmo synapse start memory --namespace=demo

Clarification & permission

A Neuron can pause and ask instead of returning a result, by returning a marker. The Axon turns it into a CLARIFICATION or PERMISSION signal; an answering Dendrite (a central Cortex or any peer) replies. No extra client - the Engram is the memory, the return-marker is the request.

async def writer(input, context, *, recall, imprint):
    # 1. Try the Engram of standing grants first.
    if input.get("permission"):                      # resumed with a verdict
        if input["permission"]["granted"]:
            await imprint("grants", op="upsert", merge_key="action",
                          entry={"action": "write_file", "granted": True})
            return {"wrote": True}
        return {"wrote": False}
    hit = await recall("grants", query={"action": "write_file"})
    if hit.hits:
        return {"wrote": True}                        # already allowed
    # 2. Miss -> ask. Same shape as {"__clarification__": True, ...}.
    return {"__permission__": True, "action": "write_file",
            "scope": {"path": "/tmp/out"}, "reason": "persist output"}

# Answering side (centralised or decentralised  -  just who subscribes):
@cortex.on_permission
async def decide(sig):
    granted = sig.payload["action"] == "write_file"
    # Re-dispatch a TASK carrying the verdict so the Neuron resumes:
    await cortex.respond_to_permission(sig, granted=granted, reason="allowlisted")
    # ...or emit a discrete signal instead:
    # await cortex.grant_permission(sig, reason="allowlisted", ttl_ms=3_600_000)

on_clarification / respond_to_clarification / answer_clarification are the clarification-side equivalents. See design/ENVELOPE_SPEC.md §7.5.

Provider-backed Neurons

Drop an LLM into any workflow with Neuron(source=...):

from cosmonapse import Axon, Neuron

# Local Ollama daemon
axon = Axon(
    neuron_id="chat",
    neuron_fn=Neuron(source="ollama", model="llama3"),
)

# HuggingFace TGI / vLLM / llama.cpp / LM Studio
axon = Axon(
    neuron_id="summariser",
    neuron_fn=Neuron(
        source="huggingface",
        endpoint="http://localhost:8080",
    ),
)

# HuggingFace hosted endpoint (requires api_key)
axon = Axon(
    neuron_id="classifier",
    neuron_fn=Neuron(
        source="huggingface",
        endpoint="https://<your-endpoint>.endpoints.huggingface.cloud",
        api_key="hf_…",
        use_chat_api=True,
    ),
)

# OpenAI (api_key falls back to OPENAI_API_KEY)
axon = Axon(
    neuron_id="writer",
    neuron_fn=Neuron(source="openai", model="gpt-4o", system="Be concise."),
)

# Anthropic (api_key falls back to ANTHROPIC_API_KEY)
axon = Axon(
    neuron_id="reasoner",
    neuron_fn=Neuron(source="anthropic", model="claude-opus-4-6"),
)

# OpenAI-compatible hosted endpoints  -  groq, openrouter, together, mistral.
# Pre-configured huggingface Neurons; api_key falls back to <PROVIDER>_API_KEY.
axon = Axon(
    neuron_id="fast",
    neuron_fn=Neuron(source="groq", model="llama-3.1-70b-versatile"),
)

Neuron(...) returns an async callable satisfying NeuronFn. Pass prompt or messages (OpenAI-style) in the task input. Output is always {"response": "<text>", "meta": <raw provider payload>}.

Requires httpx (pip install httpx).

Synapse adapters

URL factory:

from cosmonapse import connect_synapse, synapse_from_url

synapse = await connect_synapse("cosmo://127.0.0.1:7070")   # DevSynapse
synapse = await connect_synapse("nats://nats:4222")          # NatsSynapse
synapse = await connect_synapse("kafka://broker:9092")       # KafkaSynapse

For MemorySynapse, construct directly:

from cosmonapse import MemorySynapse
synapse = MemorySynapse()
await synapse.connect()

Storage backends

A RegistryStore is the orchestrator's view of the live Neuron population, rebuilt from REGISTER / HEARTBEAT / DEREGISTER Signals. Pass one as registry_store= when an orchestrator dispatches by capability instead of by neuron_id.

Pathway

dendrite.dispatch(...) returns a Pathway - a per-trace event handle with three consumption shapes on the same primitive:

# 1. Sequential request/reply - block for the first terminal Signal.
sig = await orch.dispatch_and_wait(neuron="answerer", input={"q": "hi"}, timeout_s=5.0)
print(sig.payload["output"])

# 2. Reactive - trace-scoped callbacks.
pw = await orch.dispatch_and_subscribe(capabilities=["plan"], input={"goal": "..."})

@pw.on(SignalType.THOUGHT_DELTA)
async def stream(s):
    print(s.payload["delta"], end="")

# 3. Streaming iteration.
async with await orch.dispatch(capabilities=["plan"], input={"goal": "..."}) as pw:
    async for sig in pw:
        if sig.type is SignalType.AGENT_OUTPUT:
            break

Pathway(scope=...) filters delivery. "all" (default) sees every PATHWAY_TYPES Signal on the trace; "terminal" delivers only FINAL / ERROR / CLARIFICATION / PERMISSION - the decentralised pattern where intermediate orchestration is handled peer-to-peer and the Cortex only wakes for a conclusion or a blocked decision. Pathways auto-close on FINAL or ERROR. observe_pathway(trace_id) opens a Pathway in observer role for a trace another peer started. See the 04-pathway example.

Engram (shared memory)

An Engram is a memory backend addressed over the bus. The Neuron stays pure - it gains keyword-only recall / imprint parameters that the Axon injects because the Axon was constructed with engrams=[EngramBinding(...)]. Under the hood recall(...) emits a RECALL Signal on the current trace and awaits the matching RECALLED reply.

from cosmonapse import Axon, Dendrite, EngramBinding, InMemoryEngram

async def researcher(input, context, *, recall, imprint):
    prior = await recall("ctx", query={"text": input["question"]})
    if prior.hits:
        return {"answer": prior.hits[0].entry["answer"], "source": "cache"}
    answer = solve(input["question"])
    await imprint("ctx", op="upsert", merge_key=f"q:{input['question']}",
                  entry={"question": input["question"], "answer": answer})
    return {"answer": answer, "source": "computed"}

# Host owns the backend; engram_id is the wire address.
host.attach_engram(InMemoryEngram(engram_id="ctx", engram_kind="context"))

# Worker binds a local name to that wire id, so ops can repoint the
# backend without editing Neuron code.
worker.attach_axon(Axon(
    neuron_id="researcher", neuron_fn=researcher, capabilities=["research"],
    engrams=[EngramBinding(name="ctx", directed_id="ctx")],
))

Backends: InMemoryEngram (tests), SqliteEngram (single-host durable), PostgresEngram (pip install cosmonapse). See the 06-engram-integration and 11-rag examples.

The cosmo CLI

pip install cosmonapse also installs the cosmo command:

Watch any workflow live:

cosmo synapse start memory --namespace=demo
cosmo doppler --prism --url=cosmo://127.0.0.1:7070 -n demo   # opens http://127.0.0.1:7071

Examples

The companion cosmonapse-examples repo has 13 runnable, numbered examples - from a single-process hello-world up through capability routing, competitive bidding, a full hybrid-retrieval RAG system, an MCP-backed coding agent, and retry / STOP / saga-rollback. Start with 01-quickstart.

License

MIT (c) 2026 Aqib Khan