zeroclaw-api-client-python
Your ZeroClaw, from Python. Typed calls, live streams, resilient by default.
Docs · Quick start · Guides · Resilience · Endpoint map · Architecture · ZeroClaw
zeroclaw-api-client is a typed, resilient, async Python client for the
ZeroClaw gateway: its HTTP API,
its Server-Sent Events stream and its WebSocket chat. It is the Python port
of zeroclaw-api-client-rust
and zeroclaw-api-client-go,
with the same surface, behavior and tests.
An independent, community-maintained client. It is not affiliated with or endorsed by ZeroClaw Labs; ZeroClaw is a trademark of ZeroClaw Labs.
pip install zeroclaw-api-client # or: uv add zeroclaw-api-client
import asyncio
from zeroclaw_client import Client
async def main() -> None:
async with Client("http://127.0.0.1:42617", token="zc_…") as client:
status = await client.system.status()
channels = await client.channels.list()
await client.config.set("agents.helper.model", "gpt-5.4-mini")
async with await client.chat("helper") as chat:
reply = await chat.ask("How many orders are pending?")
print(status.version, len(channels), reply.full_response)
asyncio.run(main())
Why an SDK
ZeroClaw is a self-hosted AI agent runtime. Everything an operator can do with it (configure agents, models and channels, pair devices, run SOPs, schedule jobs, read sessions and costs, talk to an agent) is exposed by its gateway over HTTP and WebSockets. Calling that surface by hand from every application that embeds ZeroClaw means repeating, in each of them:
- the contract: about 150 routes, a dozen response shapes for errors, a JSON Patch dialect for the configuration, an event stream whose frames have no names, a chat protocol with typed frames;
- the edge cases: secrets that are never returned, paths that must be encoded segment by segment, a gateway that drops connections while it reloads, routes that only exist with some build features;
- the resilience: timeouts that must not cut a stream, retries that must never repeat a write, back-pressure.
This package does it once, typed and tested, so applications talk to ZeroClaw in terms of channels, agents and sessions instead of URLs and JSON. When the gateway changes, one place changes.
Features
-
The whole client surface of the gateway, one attribute per area:
Attribute Covers configper-field reads and writes, atomic JSON Patch, map aliases (agents, providers, channels), drift, catalogs, setup sections pairingpairing codes, bearer tokens, paired devices channelschannel state, sending, group chats and their members, identity binding, relinking chat()agent chat over WebSocket: streamed replies, tool calls, tool approvals sessions,eventstranscripts, session state, the live Server-Sent Events stream agent(),files,skillsagent workspaces, personality files, shared files, image uploads, skill bundles cron,memory,costscheduled jobs, long-term memory, usage and cost sops,canvas,quickstartSOP authoring, runs and approvals; canvases; the first-agent setup systemhealth, status, metrics, diagnostics, tools, logs, upgrades, reload webhook,webauthnone-shot webhook and A2A messages; hardware-key ceremonies -
Async from end to end (
asyncio): one client is shared by every task; the streams are async iterators. -
Resilient by default, built as a stack of layers over an
httpxtransport: answer timeouts that never cut a stream, retries with jittered exponential backoff for idempotent calls only that honorRetry-After(tenacity), and, opt-in, a circuit breaker (purgatory), a client-side rate limit (aiolimiter) and a shared concurrency limit; plus your own middleware (metrics, tracing, signing) around every request. -
Typed end to end: every answer is a pydantic model validated on arrival, the package ships
py.typedand passesmypy --strict. -
Typed errors: every failure is a
ZeroClawError; a gateway refusal is anAPIErrorthat keeps its status, its explanation and its JSON body;is_not_found,is_unauthorized,is_rate_limitedandis_transientclassify any error. -
Pull, not callbacks: the event feed is an
EventStream, the canvas and SOP-run feeds aFeed, and a chat a session you iterate withasync for. The client starts no background work of its own. -
Safe by construction: every path segment is percent-encoded, the token never appears in
repr, nothing is printed or logged.
Technologies
| Concern | Choice |
|---|---|
| Language | Python 3.11+ (asyncio) |
| HTTP | httpx (HTTP/1.1, optional HTTP/2, proxies, TLS) |
| Middleware | layered transport: deadline, rate limit, retry (jitter, Retry-After), concurrency limit, circuit breaker, your own Middleware |
| Retries | tenacity |
| Rate limiting | aiolimiter leaky bucket |
| Circuit breaking | purgatory |
| WebSockets | websockets (zeroclaw.v1 subprotocol, keep-alive pings) |
| Models | pydantic v2 |
| Tests | pytest with an httpx.MockTransport gateway, an in-test WebSocket gateway, captured response vectors, parametrized tests |
| Quality | ruff (format and lint), mypy strict |
| Docs | Google-style docstrings and an mdBook book |
| Supply chain | pip-audit, a locked uv.lock, PyPI trusted publishing |
Architecture
your code
│ await client.channels.rooms("whatsapp.main")
▼
Client ── resource attributes (config, channels, chat, …) ── pydantic requests/responses
│
├─ HTTP ──► your middleware ─► circuit breaker ─► concurrency limit ─► retry ─► rate limit ─► deadline ─► httpx transport ─► gateway
├─ SSE ───► the same stack; the body is read as a stream of events
└─ WebSocket ─► websockets (chat, canvas and SOP-run feeds)
Each method's docstring names the route it calls; the endpoint map is generated from them.
Resilience
from zeroclaw_client import CircuitBreaker, Client, RateLimit, RetryPolicy
client = Client(
"http://127.0.0.1:42617",
token="zc_…",
timeout=10, # wait for an answer at most 10 s
retry=RetryPolicy(max_retries=5, base_delay=0.25, max_delay=8),
concurrency_limit=4, # shared by every copy
rate_limit=RateLimit(rate=10, burst=20), # 10 requests/s, bursts of 20
circuit_breaker=CircuitBreaker(), # open after 5 failures in a row, for 30 s
middleware=[my_middleware], # async (request, next) -> response
)
Retries apply to GET, HEAD, OPTIONS, PUT and DELETE after a
connection failure, a connection dropped before the answer, a timeout, or a
408, 429, 502, 503 or 504, with jittered backoff and the gateway's
Retry-After honored. A POST or PATCH is never repeated. An open
breaker fails calls at once with CircuitOpenError; the rate limit fails a
wait longer than the timeout with RateLimitedError. The timeout bounds the
wait for the answer's headers, so the event stream stays open.
Examples
Runnable scripts under examples/:
| Example | Shows |
|---|---|
status.py |
version, model and channels |
pairing.py |
exchanging a pairing code for a token |
config.py |
reading fields, listing agents, an atomic patch |
group_chats.py |
channels, their group chats and members |
events.py |
following the live event stream, reconnecting |
chat.py |
streaming an agent's reply over the WebSocket |
announce.py |
posting on a channel through a one-off cron job |
resilient.py |
timeouts, retries, rate limit, circuit breaker, concurrency limit and a custom middleware |
ZEROCLAW_URL=http://127.0.0.1:42617 ZEROCLAW_TOKEN=zc_… python examples/status.py
Testing
- Parametrized unit tests for every piece of logic: error mapping, URL encoding, SSE parsing, chat frames, retry decisions, backoff, patch operations, schedules.
- HTTP-boundary tests for every endpoint against a mock gateway: method, path, query, body and decoding; plus retry, timeout, circuit breaker, rate limit, concurrency and middleware behavior.
- Response vectors: answers captured from a real gateway
(
scripts/capture-vectors.py, anonymized) must decode into the SDK's models and re-encode without losing a field. - Live tests against a running gateway (
pytest -m live): every read decodes real answers, and a write suite covers the rest of the surface, cleaning up after itself.
uv sync && uv run pytest
Compatibility
Targets the ZeroClaw 0.8.5 gateway; see Compatibility for the builds it is verified against and the routes that need a newer one.
Contributing
Work happens on branches with pull requests to main; see
CONTRIBUTING.md
and AGENTS.md.
License
Licensed under either of Apache License, Version 2.0 or MIT license, at your option.
The Python logo is a trademark of the Python Software Foundation, used here to say the client is written in Python, in line with the PSF Trademark Usage Policy.
Metadata
Release files for zeroclaw-api-client 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| zeroclaw_api_client-0.1.0.tar.gz | 117.9 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| zeroclaw_api_client-0.1.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 184.8 kB
Release files / zeroclaw_api_client-0.1.0.tar.gz
| Download URL | zeroclaw_api_client-0.1.0.tar.gz |
|---|---|
| Size | 117.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
3ca3bc49397da48dfd83b467ed9ed8034025475929dc7eb4d594193e8c73b380
|
|
BLAKE2b-256 checksum How to use checksums |
0c71e07ed3f4244239c5cd72e58f80a84c8c2c20362d5b1bb8594bb72a1fb3ea
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 14, 2026.
Transparency logRelease files / zeroclaw_api_client-0.1.0-py3-none-any.whl
| Download URL | zeroclaw_api_client-0.1.0-py3-none-any.whl |
|---|---|
| Size | 67.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
c1eede30719bd4e22f1d3e71b6917c30bb7bed9455168bf6fca620513da68859
|
|
BLAKE2b-256 checksum How to use checksums |
379d0f7a0b80e71be84a778107bef013e29759798bc09e49569faf806bea20d3
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 14, 2026.
Transparency log