Python SDK for the Agent Runtime Control Protocol (ARCP) — submit, observe, and control long-running agent jobs from Python.
Specification · Concepts · Install · Quick start · Guides · API reference
arcp is the Python reference implementation of ARCP, the Agent Runtime Control Protocol. It covers both sides of the wire — arcp.client.ARCPClient for submitting and observing jobs, arcp.runtime.ARCPRuntime for hosting agents — so either side can talk to any conformant peer in any language without hand-rolling the envelope, sequencing, or lease enforcement.
ARCP itself is a transport-agnostic wire protocol for long-running AI agent jobs. It owns the parts of agent infrastructure that don't change between products — sessions, durable event streams, capability leases, budgets, resume — and stays out of the parts that do. ARCP wraps the agent function; it does not define how agents are built, how tools are exposed (that's MCP), or how telemetry is exported (that's OpenTelemetry).
Requires Python 3.11 or later. The SDK ships as a single wheel containing the client, runtime, transports, ASGI/aiohttp middleware, the OpenTelemetry middleware, and the arcp CLI. Install from PyPI with pip, uv, or any PEP 517 resolver; the optional jwks extra pulls in httpx for remote JWKS verification.
pip install arcp
# or, with uv:
uv add arcp
# with the JWKS extra:
pip install "arcp[jwks]"Connect to a runtime, submit a job, stream its events to completion:
import asyncio
import contextlib
import os
from arcp import ClientInfo, WebSocketTransport
from arcp.client import ARCPClient
async def main() -> None:
client = ARCPClient(
client=ClientInfo(name="quickstart", version="1.0.0"),
token=os.environ["ARCP_TOKEN"],
)
async with contextlib.aclosing(client):
transport = await WebSocketTransport.connect("wss://runtime.example.com/arcp")
await client.connect(transport)
handle = await client.submit(
agent="data-analyzer",
input={"dataset": "s3://example/sales.csv"},
lease_request={"net.fetch": ["s3://example/**"]},
)
async for event in handle.events():
print(f"[{event['kind']}]", event["body"])
result = await handle.done
print("final:", result.final_status, result.result)
asyncio.run(main())This is the whole shape of the SDK: open a session, submit work, consume an ordered event stream, get a terminal result or error. Everything below is detail on those four moves.
ARCP organizes everything around four concerns — identity, durability, authority, and observability — expressed through five core objects:
- Session — a connection between a client and a runtime. A session carries identity (a bearer token), negotiates a feature set in a
hello/welcomehandshake, and is resumable: if the transport drops, you reconnect with a resume token and the runtime replays buffered events. Jobs outlive the session that started them. See §6. - Job — one unit of agent work submitted into a session. A job has an identity, an optional idempotency key, a resolved agent version, and a lifecycle that ends in exactly one terminal state:
success,error,cancelled, ortimed_out. See §7. - Event — the ordered, session-scoped stream a job emits: logs, thoughts, tool calls and results, status, metrics, artifact references, progress, and streamed result chunks. Events carry strictly monotonic sequence numbers so the stream survives reconnects gap-free. See §8.
- Lease — the authority a job runs under, expressed as capability grants (
fs.read,fs.write,net.fetch,tool.call,agent.delegate,cost.budget,model.use). The runtime enforces the lease at every operation boundary; a job can never act outside it. Leases may carry a budget and an expiry, and may be subset and handed to sub-agents via delegation. See §9. - Subscription — read-only attachment to a job started elsewhere (e.g. a dashboard watching a job a CLI submitted). A subscriber observes the live event stream but cannot cancel or mutate the job. Distinct from resume, which continues the original session and carries cancel authority. See §7.6.
The SDK models each of these as first-class objects; the rest of this README shows how.
Open a session, negotiate features, and reconnect transparently after a transport drop using the resume token — jobs keep running server-side while you're gone.
import asyncio
import contextlib
import os
from arcp import ClientInfo, SessionResume, WebSocketTransport
from arcp.client import ARCPClient
URL = "wss://runtime.example.com/arcp"
TOKEN = os.environ["ARCP_TOKEN"]
def new_client() -> ARCPClient:
return ARCPClient(
client=ClientInfo(name="resumable", version="1.0.0"),
token=TOKEN,
)
async def main() -> None:
first = new_client()
transport1 = await WebSocketTransport.connect(URL)
welcome = await first.connect(transport1)
session_id = welcome.session_id
resume_token = welcome.resume_token
handle = await first.submit(agent="long-running", input={})
async for _ in handle.events():
if first.latest_event_seq >= 2:
break
last_seq = first.latest_event_seq
# Drop the transport without sending session.bye; the job keeps running.
await transport1.close()
second = new_client()
async with contextlib.aclosing(second):
transport2 = await WebSocketTransport.connect(URL)
await second.resume(
transport2,
resume=SessionResume(
session_id=session_id,
resume_token=resume_token,
last_event_seq=last_seq,
),
)
# The runtime replays every event with seq > last_seq, then resumes live streaming.
asyncio.run(main())Submit a job with an agent (optionally version-pinned as name@version), an input, and an optional lease request, idempotency key, and runtime limit.
from datetime import UTC, datetime, timedelta
from arcp import LeaseConstraints
expires_at = (datetime.now(UTC) + timedelta(minutes=1)).isoformat().replace("+00:00", "Z")
handle = await client.submit(
agent="weekly-report@2.1.0",
input={"week": "2026-W19"},
lease_request={"net.fetch": ["s3://reports/**"]},
lease_constraints=LeaseConstraints(expires_at=expires_at),
idempotency_key="weekly-report-2026-W19",
max_runtime_sec=300,
)
print("job_id =", handle.job_id)
print("effective lease =", handle.lease)
print("resolved agent =", handle.agent_ref)Iterate the ordered event stream — log, thought, tool_call, tool_result, status, metric, artifact_ref, progress, result_chunk — and optionally acknowledge progress so the runtime can release buffered events early.
from arcp import ClientInfo
from arcp.client import ARCPClient, AutoAckOptions
client = ARCPClient(
client=ClientInfo(name="ack-demo", version="1.0.0"),
token=TOKEN,
# Coalesced session.ack: send when 32+ events have accrued, at most every 250 ms.
auto_ack=AutoAckOptions(every_n=32, interval_sec=0.25),
)
handle = await client.submit(agent="chatty", input={})
async for event in handle.events():
kind = event["kind"]
body = event["body"]
if kind == "log":
print(body.get("message"))
elif kind == "tool_call":
print("->", body.get("name"), body.get("arguments"))
elif kind == "metric":
print("metric", body)
elif kind == "progress":
print("progress", body)
# Or ack manually: await client.ack(client.latest_event_seq)Request capabilities, a budget, and an expiry; read budget-remaining metrics as they arrive; handle the runtime's enforcement decisions.
from datetime import UTC, datetime, timedelta
from arcp import BudgetExhaustedError, LeaseConstraints, LeaseExpiredError
expires_at = (datetime.now(UTC) + timedelta(minutes=10)).isoformat().replace("+00:00", "Z")
handle = await client.submit(
agent="web-research",
input={"iterations": 8, "per_call_usd": 0.3},
lease_request={
"tool.call": ["search.*", "fetch.*"],
"cost.budget": ["USD:1.00"],
},
lease_constraints=LeaseConstraints(expires_at=expires_at),
)
print("initial budget =", handle.budget)
try:
async for event in handle.events():
if event["kind"] == "metric" and event["body"].get("name") == "cost.budget.remaining":
body = event["body"]
print(f"budget remaining: {body['value']:.2f} {body.get('unit', '')}")
await handle.done
except (BudgetExhaustedError, LeaseExpiredError) as err:
# Never retryable: resubmit with a fresh lease or budget instead.
print("job ended:", err.code, err.message)Attach read-only to a job submitted elsewhere and observe its live stream (with optional history replay) without cancel authority.
from arcp import ClientInfo, ListJobsFilter, WebSocketTransport
from arcp.client import ARCPClient
observer = ARCPClient(
client=ClientInfo(name="dashboard", version="1.0.0"),
token=TOKEN,
features=("list_jobs", "subscribe"),
)
await observer.connect(await WebSocketTransport.connect(URL))
listing = await observer.list_jobs(filter=ListJobsFilter(status=("running",)))
sub = await observer.subscribe(listing.jobs[0].job_id, history=True)
print(f"subscribed from seq={sub.subscribed_from} replayed={sub.replayed}")
async for event in sub.handle.events():
print(f"[seq>{sub.subscribed_from}] {event['kind']}")
# ... later ...
await observer.unsubscribe(sub.job_id)Catch the typed error taxonomy and respect the retryable flag — LEASE_EXPIRED and BUDGET_EXHAUSTED are never retryable; a naive retry fails identically.
from arcp import ARCPError, BudgetExhaustedError, LeaseExpiredError
try:
handle = await client.submit(agent="flaky", input={})
await handle.done
except (LeaseExpiredError, BudgetExhaustedError):
# Resubmit with a fresh lease / budget instead of retrying.
raise
except ARCPError as err:
if err.retryable:
# Safe to retry with backoff (e.g. INTERNAL_ERROR, TIMEOUT).
...
else:
raiseARCP features this SDK negotiates during the hello/welcome handshake:
| Feature flag | Status |
|---|---|
heartbeat |
Supported |
ack |
Supported |
list_jobs |
Supported |
subscribe |
Supported |
lease_expires_at |
Supported |
cost.budget |
Supported |
model.use |
Supported |
provisioned_credentials |
Supported |
progress |
Supported |
result_chunk |
Supported |
agent_versions |
Supported |
ARCP is transport-agnostic. This SDK ships a WebSocket transport (default), an stdio transport for in-process child runtimes, and an in-memory transport for tests. WebSocket is the default for networked runtimes; stdio is used for in-process child runtimes. Select one by constructing the corresponding object (WebSocketTransport.connect(url), StdioTransport(...), pair_memory_transports()) and passing it to client.connect(transport); the arcp.middleware.asgi and arcp.middleware.aiohttp packages attach the WebSocket upgrade to Starlette, FastAPI, Litestar, Quart, or aiohttp.web servers.
Full API reference — every type, method, and event payload — is in docs/.
This SDK speaks ARCP v1.1 (draft). The SDK follows semantic versioning independently of the protocol; the protocol version it negotiates is shown above and in session.hello. A runtime advertising a different ARCP MAJOR is not guaranteed compatible. Feature mismatches degrade gracefully: the effective feature set is the intersection of what the client and runtime advertise, and the SDK will not use a feature outside it.
See CONTRIBUTING.md. Protocol questions and proposed changes belong in the spec repository; SDK bugs and feature requests belong here.
Apache-2.0 — see LICENSE.