comfy-python-sdk

The Python client for the Comfy API v2.
Submit a workflow, stream its progress, get your outputs — against self-hosted ComfyUI, Comfy Cloud, or serverless.

---

Python SDK for running ComfyUI workflows via the Comfy API v2. The same code runs against Comfy Cloud, a serverless deployment, or a self-hosted ComfyUI instance — only the COMFY_BASE_URL environment variable and an optional API key change.

Requirements and install

Requires Python 3.10+. Dependencies: httpx, blake3, pydantic (v2).

pip install comfy-sdk

To install from source instead (for local development, or to track an unreleased commit):

git clone https://github.com/Comfy-Org/comfy-python-sdk
cd comfy-python-sdk
pip install -e .

# To install everything needed to lint/type-check/test locally
pip install -e ".[dev]"

Optional dependencies

Install the optional pil extra with pip install -e ".[pil]" to use Preview.to_pil() for decoding an in-progress output preview to a PIL.Image.

For local

The SDK works against a ComfyUI instance with Comfy API v2. Comfy Cloud and serverless instances deployed from our developer platform already use Comfy API v2. For local or self-hosted instances, Comfy API v2 can be setup using the comfy-api-proxy.

Getting started

from comfy_sdk import Comfy

client = Comfy(api_key="comfyui-...")   # Comfy Cloud

wf = client.workflows.from_file("workflow_api.json")

# Input assets are hashed locally with blake3
# If the server already has an identical copy we reuse it, if not we upload the asset
# The workflow is updated with the core/ASSET reference instead of a local file path
asset = client.assets.from_file("photo.png")
wf.set_input("10", "image", asset)

# Run workflow
# Get outputs using the output node Id as a reference
job = client.run(wf)
for output in job.get_outputs("9"):
    output.to_file(output.name)

Authentication — one client, per-surface key

Surface	api_key
Comfy Cloud (https://cloud.comfy.org) — the default	Required
Serverless deployment	Required
Self-hosted ComfyUI (behind the API proxy)	Omit — no key is sent, even implicitly

client = Comfy(api_key="comfyui-...")   # Comfy Cloud

AsyncComfy takes the same arguments. A key is only ever attached to requests aimed at the target deployment's own origin — a server-returned follow-up link (job.urls.self/cancel/events, or a redirected asset download) pointing anywhere else never receives it.

Targeting another deployment

Comfy() points at Comfy Cloud and takes no base-URL argument. To run against a serverless deployment or a self-hosted instance behind comfy-api-proxy, set COMFY_BASE_URL in the environment:

export COMFY_BASE_URL="https://<deployment>.run.comfy.app"  # serverless
export COMFY_BASE_URL="http://127.0.0.1:8189"               # self-hosted proxy

It is read each time a client is constructed, must be an http(s) URL, and an unset or blank value (including whitespace-only) means Comfy Cloud.

Upgrading from an earlier version: Comfy("<url>", "<key>") becomes Comfy(api_key="<key>") with COMFY_BASE_URL set. api_key is keyword-only, so the old positional call raises TypeError rather than reading a URL as a key.

The SDK identifies itself via a User-Agent header (for support and usage analytics) — this is request metadata only; no other data is collected. Pass client_info="my-app" to append an app/my-app token so an integration can attribute its own traffic:

client = Comfy(api_key="comfyui-...", client_info="my-app")

Partner (API) node auth

Workflows that use partner/API nodes (Gemini, etc.) need a Comfy API key to authenticate them. Pass it per submit with api_key=. This is not the same as the api_key you construct Comfy with: the constructor key authenticates you to the server, while this one authenticates the partner nodes inside the workflow (it is often the same comfyui-… key):

job = client.run(wf, api_key="comfyui-...")
# or drive it yourself:
job = client.submit(wf, api_key="comfyui-...")

The SDK sends it once as extra_data.api_key_comfy_org alongside the workflow — one key authenticates every partner node in the graph. It is never logged or persisted by the SDK. Omit api_key and no extra_data is sent at all.

Assets and core/ASSET

client.assets.from_file(...) / from_bytes(...) / from_stream(...) / from_url(...) return a lazy asset handle immediately — no network call yet. Embed it directly into the workflow graph:

asset = client.assets.from_file("photo.png")
wf.set_input("10", "image", asset)

On first use (submitting the workflow, or an explicit asset.commit()), the SDK:

1. hashes the bytes locally with blake3;
2. probes the server's dedup fast-path — a HEAD existence check by hash, then a cheap from-hash mint if the server already has those bytes;
3. only streams a full multipart upload on a miss.

At submit time, every asset handle found anywhere in the graph is replaced by a core/ASSET reference object ({"__type": "core/ASSET", "info": {"id": ..., "hash": ..., "file_path": ...}}), which the server resolves back to the uploaded asset when it runs the workflow.

Once committed, an asset also carries job_id — the id of the job that produced it, or None for an asset with no producing job (e.g. a plain upload) — and expires_at, its retention deadline, or None if it doesn't expire.

Delete an asset with asset.delete(), or by id alone with client.assets.delete(asset_id):

asset.delete()
# or, without holding a handle:
client.assets.delete(asset_id)

Deletion needs a proxy new enough to serve DELETE /api/v2/assets/{id} — an older comfy-api-proxy returns 405 instead. (AsyncAsset.delete() / AsyncAssetFactory.delete() mirror both with await.)

Live progress

job = client.submit(wf)
for event in job.events():          # SSE; live, auto-reconnecting (no replay)
    match event:
        case Progress() as p:       print(f"{p.value:.0%} {p.message}")
        case Preview() as pv:       show(pv.to_pil())
        case OutputReady() as o:    o.output.to_file(f"partial/{o.output.name}")
        case StatusChange(status="succeeded"): break
result = job.result()               # raises JobFailed with node details on failure

job.events() reconnects automatically if the stream drops, but never replays a frame you've already seen (the stream carries no cursor). That's why polling stays authoritative: job.wait() / job.result() (and client.run(), which is submit() + result()) always fall back to GET /jobs/{id} to decide when a job is really done — use events() for live UI feedback, and wait()/result()/run() for the definitive answer. job.status is the current status string; job.outputs is the full list of output handles regardless of which node produced them (job.get_outputs(node_id) filters to one node, as in the quickstart above).

Getting a job's workflow back

The SDK only holds the workflow it submitted for as long as the originating Job handle stays alive — for a job rehydrated purely by id (client.jobs.get(job_id)), get_workflow() is the only way to see the graph:

job = client.jobs.get(job_id)
wf = job.get_workflow()
match wf.format:
    case "api":   ...   # the executed graph; frontend-only nodes already resolved away
    case "save":  ...   # the authoring workflow at the pinned version, canvas layout intact

format discriminates the shape of wf.graph, so branch on it rather than assume one. It depends on how the job was submitted, not on anything a caller controls — jobs submitted through this SDK always get "api" today, since v2 submission has no version-pinning fields yet. (AsyncJob.get_workflow() mirrors this with await.)

Downloading outputs

A finished job exposes its results as Output handles — job.outputs, or job.get_outputs(node_id) to filter to one node. Each output is an asset you can pull down whichever way suits the caller:

out = job.get_outputs("13")[0]
out.to_file("result.png")                   # stream to disk in chunks
data = out.to_bytes()                       # buffer into memory
out.to_file("head.png", range=(0, 1023))    # range-aware: first 1 KiB only

Every output also carries job_id, the id of the job that produced it — so a caller holding just an output can get back to the job that made it.

get_download_url() hands back a fetchable URL instead of transferring the bytes through your process — give it to a browser, a CDN, or another service:

link = out.get_download_url()               # DownloadUrl(url=..., expires_at=...)

On Comfy Cloud / serverless the URL is a short-lived, self-authorizing signed storage URL: whoever holds it can read the asset until expires_at with no API key of their own. On a self-hosted proxy it's the content endpoint (normal auth still applies) and expires_at is None. It works on every backend and never downloads the bytes first. (AsyncOutput mirrors all of the above with await.)

Sync and async

Comfy and AsyncComfy expose the identical surface — swap the import and add await / async for:

from comfy_sdk import AsyncComfy

async def main() -> None:
    async with AsyncComfy(api_key="comfyui-...") as client:
        wf = client.workflows.from_file("workflow_api.json")
        job = await client.run(wf)
        await job.outputs[0].to_file("out.png")

Typed errors

comfy_sdk translates the API's error envelope into a small set of exceptions, all importable from the top-level package and all subclasses of ComfyError:

• Unauthorized, Forbidden, NotFound — auth and lookup failures.
• InvalidWorkflow, WorkflowFormatUi — the graph itself was rejected; WorkflowFormatUi specifically means a UI-export (nodes/links/ last_node_id) was submitted instead of the API-format graph — the SDK catches this locally before it ever reaches the server.
• MissingAsset — a core/ASSET reference could not be resolved.
• HashMismatch, BlobNotFound — asset upload/dedup failures.
• IdempotencyKeyReuse — the Idempotency-Key was reused. submit() (and run()) attach a fresh key to every call, so an accidental exact resend never runs the workflow twice. Keys are single-use — reject-on-duplicate, there is no replay — so if you pass your own idempotency_key= and reuse it, the second call raises this. After an ambiguous failure (e.g. a timeout where you don't know if the job was created), poll or list your jobs rather than resubmitting with the same key.
• InsufficientCredits — the account can't afford the job.
• QueueFull — backpressure; carries .retry_after seconds. client.submit already retries this automatically for a bounded budget before giving up and raising it.
• JobFailed — a job reached a non-succeeded terminal state; .error carries node-level detail when the platform provided one.

from comfy_sdk import JobFailed, QueueFull, Unauthorized

try:
    result = client.run(wf)
except JobFailed as e:
    print(e.error)
except Unauthorized:
    print("check your api_key")

Architecture — two layers

• comfy_low — generated protocol bindings. Pydantic v2 models generated from spec/openapi.yaml (src/comfy_low/models/_generated.py, committed; regenerate with scripts/gen_models.sh, CI fails on drift) plus a thin hand-written httpx transport (sync + async), one function per operationId, with the mandatory escape hatches: raw response access, unbuffered/streaming bodies, all headers, and per-request timeout/abort. Boring and replaceable.

• comfy_sdk — the idiomatic layer integrators import. This is where the value lives: blake3 content-addressed dedup-upload, core/ASSET substitution, idempotent submit, live SSE with reconnect, poll-authoritative run(), range-aware downloads, and typed exceptions mapping the error envelope.

spec/openapi.yaml is a one-way vendored copy of the canonical Comfy API v2 contract — do not hand-edit it (see spec/README.md). It's synced periodically from that canonical contract, stripped of anything tagged internal, and pinned by spec/VERSION.

Related projects

Clients for the same Comfy API v2 contract:

Project	Language	Package
comfy-python-sdk	Python	comfy-sdk
comfy-typescript-sdk	TypeScript	@comfyorg/sdk

Development

pip install -e ".[dev]"
ruff check .
ruff format --check .
mypy src
pytest -v

Regenerating and checking the vendored protocol layer (a separate CI job):

pip install -e ".[codegen]"
python scripts/gen_models.sh     # regenerate comfy_low models from spec/openapi.yaml
python scripts/check_drift.py    # same check CI runs; fails if committed models drifted

Releases

Releases are published to PyPI from a GitHub Release (tag vX.Y.Z) by .github/workflows/publish.yml, using PyPI's Trusted Publishing (OIDC) — no API token is stored in this repo.