pipelex-sdk

The Python client for the Pipelex hosted API.

pipelex-sdk is the Python counterpart of @pipelex/sdk, exactly as mthds (the mthds-python package) is the Python counterpart of the mthds npm package. It is the hosted superset: the five normative MTHDS Protocol routes (inherited from mthds) plus the durable run lifecycle plus the Pipelex product surface (methods, organizations, billing, API keys, onboarding, storage, run records).

One-way dependency: pipelex-sdk → mthds.

Install

pip install pipelex-sdk

Run a method from the command line

The package publishes the pipelex-sdk command, so a method runs on the hosted API with nothing to install but uv:

export PIPELEX_API_KEY=<your key>
uvx pipelex-sdk run --method github.com/acme/methods/receipt-review@v1.0.0 --inputs-template > inputs.json
uvx pipelex-sdk run --method github.com/acme/methods/receipt-review@v1.0.0 --inputs inputs.json

--method takes a published address, a catalog id (mt_…, which runs the method's latest published version, mt_…@<n> for a fixed version or mt_…@draft for its draft) or a local .mthds file or bundle directory. The run's main output is printed as JSON on stdout, and the run id, each uploaded file and every error go to stderr; Ctrl-C leaves the run going on the server and names it. uvx pipelex-sdk script --method <address | mt_id> checks the method and writes a short shell script that runs it with this SDK's version pinned. The key comes from PIPELEX_API_KEY only, and no .env file is read. The command is the twin of npx @pipelex/sdk, held to the same recorded cases; docs/cli.md describes it.

Configuration

The API key resolves, in order: explicit api_key argument → PIPELEX_API_KEY → anonymous. The token is optional — anonymous access works against the protocol routes (e.g. a local bare runner); the product routes return 401.

The base URL resolves, in order: explicit base_url argument → PIPELEX_BASE_URL → the hosted default https://api.pipelex.com. The base URL is host-only (no path/query/fragment); every endpoint composes as {base}/v1/{endpoint}.

The SDK never reads the mthds resolver (MTHDS_API_KEY / MTHDS_BASE_URL / ~/.mthds/config) — those settings configure the vendor-neutral mthds tooling and whichever runner it targets, not this Pipelex client.

request_timeout_seconds (constructor argument, default 20 min) is the time limit of the routes that can take long: the blocking execute, validate, models, and a start carrying a bundle (mthds_contents, files or bundle_b64) or a method_ref, which the server resolves before it answers. version and any other start answer fast, so they take request_timeout_seconds capped at the 30-second poll budget, the hosted gateway's own cut-off. The SDK's own polls and product requests keep budgets of their own.

app_info (constructor argument, an AppInfo from pipelex_sdk.user_agent) puts your application's name in front of the SDK's own tokens in the User-Agent that every request carries — acme-invoicer/1.4.0 pipelex-sdk-python/0.11.0 mthds-python/0.15.0 python/3.12.4 (linux; x86_64) — which the platform uses to attribute traffic in its analytics. The header follows the spec conformance/specs/client-identification.md, in the conformance repository, where the cross-repo specs sit beside the tests that verify them; see docs/client-identification.md.

The client is async-only (httpx AsyncClient) and is an async context manager.

Quickstart

from pipelex_sdk.client import PipelexAPIClient
from pipelex_sdk.validation_models import VALIDATION_VIEW_INPUT_FORM


async def main() -> None:
    async with PipelexAPIClient() as client:
        # 1. Validate an MTHDS bundle. The verdict is always returned (never raised):
        #    a 200 discriminated on `is_valid`, carrying `rendered_markdown`.
        #    Structured views are opt-in: asking for `input_form` here is what populates
        #    `report.input_form` (the per-pipe input-form descriptors); omit `views` and the
        #    request body carries no `views` key at all.
        report = await client.validate([bundle_text], views=[VALIDATION_VIEW_INPUT_FORM])
        print(report.rendered_markdown)
        if not report.is_valid:
            return

        # 2. Run a method end-to-end. `start_and_wait` self-heals across runner kinds:
        #    durable start+poll on the hosted API, blocking execute on a bare runner.
        result = await client.start_and_wait(
            pipe_code="my_pipe",
            inputs={"topic": "quantum computing"},
        )

        # 3. Read the output. Every completed run delivers a resolved `main_stuff`
        #    (the full working memory also rides `pipe_output` on the blocking path).
        print(result.main_stuff)

Run a published method by address

A method reaches every method-taking call in exactly one of three forms: inline source, a method_ref address (github.com/<owner>/<repo>[/<selector>][@<tag>], resolved by the server — pipelex-api >= 0.21.0; on api.pipelex.com availability follows the platform deploy that forwards it), or a hosted method_id (mt_…, resolved by the platform). A method_ref pairs with nothing — it is a complete run source — and its runs carry typed provenance:

ack = await client.start(method_ref="github.com/Pipelex/methods/documents@v0.1.0", inputs={...})
print(ack.method_provenance.commit_sha)  # the SHA actually fetched — stable even if the tag moves
result = await client.wait_for_result(ack.pipeline_run_id)

# The tooling routes take the same selectors under a strict XOR (exactly one, no pairing):
report = await client.validate(method_ref="github.com/Pipelex/methods/documents@v0.1.0")
report = await client.validate(method_id="mt_123")

# A saved method's bare id runs its latest published version, `mt_…@3` version 3 and
# `mt_…@draft` its draft; the ack says which version runs.
ack = await client.start(method_id="mt_123@3", inputs={...})
print(ack.method_version)  # 3

Draft, publish and pin a saved method

A saved method has a draft, written freely and never validated, and immutable published versions numbered from 1. updated_at is the draft's token: echo it so no write or publish takes a draft you have not seen. The method routes take a bare id, and parse_method_selector strips a suffix:

from pipelex_sdk.method_selector import parse_method_selector
from pipelex_sdk.product_models import MethodDraftInput, MethodPublishRefused

method = await client.get_method("mt_123")
draft = await client.write_draft("mt_123", MethodDraftInput(mthds="domain = 'demo'\n", expected_updated_at=method.updated_at))
result = await client.publish_method("mt_123", expected_draft_updated_at=draft.updated_at)
if isinstance(result, MethodPublishRefused):
    print(result.message, result.validation)
else:
    print(f"version {result.version.version}")  # published, or unchanged
versions = await client.list_method_versions("mt_123")  # newest first
parsed = parse_method_selector("mt_123@3")  # ParsedMethodSelector(method_id="mt_123", version=3)

Read a method's inputs and outputs

pipe_io() returns a method's pipe I/O contracts, input form and output form in one call, without validating it (no dry run). The server selects the pipe — your pipe_ref, else the package manifest's main_pipe, else the closure's single main_pipe declaration — and the three artifacts are the standard's own models from mthds.protocol:

from pipelex_sdk.crate_models import PipeIORequest, PipeIOValidReport

report = await client.pipe_io(PipeIORequest(method_ref="github.com/Pipelex/methods/documents@v0.1.0"))
if isinstance(report, PipeIOValidReport) and report.pipe_ref is not None:
    descriptor = report.input_form[report.pipe_ref]
    print([field.name for field in descriptor.fields], report.is_runnable)

all_pipes=True keys the three maps by every pipe instead, and include_files=True echoes the closure's .mthds files. prepare_inputs reads its signature from this route, so it needs an API that serves POST /v1/pipe-io.

Generate typed code into your project

codegen() projects a method into stamped typed artifacts plus their codegen.lock, and write_codegen_tree writes that response to disk verbatim, so the tree is byte-identical to a local pipelex codegen types run and no pipelex install is needed:

from pathlib import Path

from pipelex_sdk.codegen_writer import write_codegen_tree
from pipelex_sdk.crate_models import CodegenRequest, CodegenValidReport

report = await client.codegen(CodegenRequest(method_ref="github.com/Pipelex/methods/documents@v0.1.0", target="python-pydantic"))
if not isinstance(report, CodegenValidReport):
    raise SystemExit(report.message)
written = write_codegen_tree(report, output_dir=Path("src/generated/documents"))
print(written.written, written.removed)

It never overwrites a file codegen does not own, rewrites only what changed, and prunes stamped artifacts that dropped out of the set. Commit the tree; do not run a formatter over it.

Gate a committed tree in CI, with no key and no pipelex

run_codegen_check is the writer's counterpart: pure hashing over the tree and its lock, so it boots no engine, reaches no network and needs no API key. Point it at the directory you generated into:

from pathlib import Path

from pipelex_sdk.codegen_check import run_codegen_check

report = run_codegen_check(root=Path("src/generated/documents"))
if not report.is_current:
    for drift in report.drifts:
        print(f"{drift.path}: {drift.category} — {drift.detail}")
    raise SystemExit(1)

The drift categories are pipelex codegen check's, and so are the sentences: an artifact edited below its stamp is hand-edited, one off the locked hash is modified, one the lock tracks and disk has lost is missing, and a stamped file the lock does not track is an orphan — the stale-artifact class a per-file stamp cannot catch alone. The two readers reach the same verdict over the same bytes apart from two deliberate divergences, both documented in docs/architecture.md: this one accepts a projection line whose axes are outside its own vocabulary, where the CLI calls such a tree hand-edited, and it refuses a Python artifact that declares a PEP 263 source encoding, where the CLI calls that one current. Regeneration stays a developer action, because it needs the engine; the check is the CI action, because it needs only hashes, so an upstream template improvement never reddens your pipeline. Whether the tree still matches what the method resolves to is a separate question the engine alone can answer — compare report.crate_fingerprint against a live codegen() response to close it.

Check a model reference

check_model_reference() asks the runner whether a reference a method's model field could name resolves, as what kind and to which model (GET /v1/models/check, served by pipelex-api from pipelex 0.78.0). A reference that resolves nowhere is a verdict, not an error, and the verdict is one arm per reference kind, so narrowing it types its matches:

from pipelex_sdk.model_reference_models import ModelCheckCategory, ModelReferenceResolution, PresetReferenceVerdict

verdict = await client.check_model_reference("$writing-factual", category=ModelCheckCategory.LLM)
if verdict.resolution is ModelReferenceResolution.RESOLVED:
    for entry in verdict.matches:
        # `resolves_to` is None when the name exists but a run through it would reach no model.
        print(entry.category, entry.resolves_to or "warning: no model the runner can call")
    if isinstance(verdict, PresetReferenceVerdict):
        print(verdict.matches[0].description)
else:
    print(verdict.other_kinds, verdict.suggestions)  # what the caller may have meant

A reference the runner cannot read at all (blank, a sigil alone, too long) or an unknown category raises ApiResponseError, a 422 whose error_type is InvalidModelReference or InvalidModelCategory.

Long runs: start + poll explicitly

Behind the hosted gateway, a synchronous execute() is cut off at ~30s and surfaces a PipelineExecuteTimeoutError pointing here. For long methods, drive the durable lifecycle yourself — the run survives client disconnects and is resumable by pipeline_run_id:

ack = await client.start(pipe_code="long_pipe", inputs={...})
result = await client.wait_for_result(ack.pipeline_run_id)

When a run is refused: ApiResponseError says why and what to do next

A method the plane will not run — a bundle naming a model the deck does not serve, a pipe whose output cannot be assembled — is refused before any result exists: start, execute and start_and_wait raise ApiResponseError on the 422. Its message names the request, gives the reason and, on its own line, the next step the plane advises, so printing the error is already actionable:

API POST /v1/start failed (422): Pipe 'draft_pitch' (PipeLLM), field 'model': Model handle 'gpt-5.1' was not found in the model deck

Did you mean: gpt-5.5, gpt-5.4, gpt-5.6-sol, gpt-5.4-pro, gpt-5.6-luna
Next step: Edit the bundle as each validation error says: apply its suggested fix where it has one, after confirming an unsafe one

A program reads the same thing as fields:

from pipelex_sdk.errors import ApiResponseError

try:
    result = await client.start_and_wait(pipe_code="pitch_product", mthds_contents=[bundle])
except ApiResponseError as exc:
    if exc.error_domain == "input":
        for item in exc.validation_errors or []:
            print(f"{item.pipe_code}: {item.message}")  # the failing pipe, typed as ValidationErrorItem
        if exc.user_action is not None:
            print(f"Next step: {exc.user_action.detail}")
    raise

validation_errors is set when the runner itemized its refusal at load; a run that failed during execution carries the pipe in server_message instead (Pipe 'analyze_topics' failed (review_topics → analyze_topics): …).

When a run fails: RunFailedError carries the run's report

A run that ends without a result — FAILED, CANCELLED, TERMINATED or TIMED_OUT — makes wait_for_result, start_and_wait and download_artifacts raise RunFailedError. Its message already names the status and the reason (Run finished with status FAILED: <message>), status is the typed RunStatus, and error is the run's stored error report, typed whole as RunErrorReport (pipelex_sdk.error_models): the runner's error_type, message, title, type_uri, error_domain, error_category, retryable, user_action, model, provider, provider_metadata, validation_errors, and anything newer on model_extra. error is None for a run that ended with no report, such as a cancelled one.

from pipelex_sdk.errors import RunFailedError

try:
    result = await client.wait_for_result(run_id)
except RunFailedError as exc:
    report = exc.error
    if report is None:
        print(f"Run {exc.run_id} ended {exc.status} without a report.")
    else:
        print(f"{report.title}: {report.user_action.detail if report.user_action else report.message}")
        if report.retryable:
            ...  # the same run may succeed if started again

Branch on error_domain (input, config, runtime), type_uri and retryable, never on the wording of message. The report is the runner's verbose one, so message and provider_metadata can hold a model provider's raw text: what a person should see of it is your application's decision. The same report is on RunRead.error when you read the run's status, on RunResultFailed.error from get_run_result, and on RunHistoryItem.error in the run lists.

API errors: branch on type_uri and error_domain, not the HTTP status

Every /v1 route raises a typed ApiResponseError on a non-2xx answer, carrying the members of the RFC 9457 problem document: the protocol routes (execute, start, validate, models, version), the run status and results reads, and the product routes — the account, methods, organization, billing, API-key, onboarding, storage and upload methods, codegen, resolve, pipe_io and check_model_reference, and the run records (list_runs, iterate_runs, get_run_detail, update_run). It is mthds's own ApiResponseError narrowed, so except mthds.runners.api.exceptions.ApiResponseError catches it too; health raises one too, and docs/architecture.md lists the error regimes. A 2xx the SDK cannot read — a body that is not JSON, such as a gateway's HTML page, or JSON that is not what the route answers — raises an ApiResponseError as well, carrying the 2xx status and the parse failure as __cause__, runtime and not retryable (docs/errors.md). The branch fields are type_uri, the problem's type, a stable URI naming the error class that every problem carries, and error_domain, the coarse class (input means the caller can fix it, config that a configuration change is needed, runtime that execution failed). error_domain and retryable are always decided: each is the member the server sent, which the runner sends on the problems it renders — a run route's refusal, and those of codegen, resolve, pipe_io and check_model_reference, which the hosted API relays from the runner — and the platform on the problems it renders itself once its deployment classifies them, and otherwise the SDK's fallback, read from the status, as on the platform's Lambda-backed routes today:

from pipelex_sdk.errors import ApiResponseError

try:
    created = await client.create_pipelex_api_key(label="ci")
    print(created.api_key)  # plaintext — returned only once
except ApiResponseError as exc:
    if exc.type_uri == "https://pipelex.com/errors/pipelex_api_key_limit_reached":
        print("Per-account key limit reached — revoke an old key first.")
    elif exc.type_uri == "https://pipelex.com/errors/validation_failed":
        print(f"Fix the request: {exc.server_message}")
    else:
        print(f"Unexpected failure, request id {exc.request_id}")
        raise

Branch on error_domain for the class — if exc.error_domain == "input": shows the caller what to fix, whatever the exact error, and exc.retryable says whether asking again can succeed (docs/errors.md). The rest of the document rides beside them: server_message (the detail), title, instance, retryable, user_action (the mthds UserAction, kept only when it has a kind and a non-empty detail), error_category, the platform's field-level errors, validation_errors for a bundle fault, and request_id for a support request, read from the body or from the X-Request-ID header. The answer itself stays reachable as plain data: status, headers (lower-case names, so exc.headers.get("retry-after") reads a 429's delay) and request_url. code (the platform's closed code, such as conflict) and error_type (the runner's exception class name) are each surface's own finer code, and a client of that surface may branch on it too: on the platform's own problems type_uri is derived from code one to one, so exc.code == MethodErrorCode.METHOD_NOT_PUBLISHED and the matching type_uri read the same answer, while a problem the platform relays from the runner carries the runner's type_uri and error_type and no code. A code the caller does not know gets its generic handling, since the platform adds codes; type_uri and error_domain stay the fields to branch on across both surfaces. problem is the decoded document whole, for any member the SDK does not name.

No answer at all: ApiUnreachableError

When a request gets no answer — a refused connection, a host name that does not resolve, a TLS failure, a timeout — every method of the client raises ApiUnreachableError, never httpx's own exception: the protocol routes (execute, start, validate, models, version) as well as the run reads, wait_for_result, start_and_wait, the product routes and health. It carries api_url, the base URL that could not be reached, and code: ABORT_TIMEOUT when the client's own time limit ran out once the request had reached the API, while it was sent or answered, and otherwise the name of the httpx failure (ConnectError, ReadError, ConnectTimeout, PoolTimeout, …), which stays reachable as __cause__. Like ApiResponseError, it is a PipelineRequestError, so one except PipelineRequestError covers every failed request. A blocking execute cut off by a timeout after about 28 seconds raises PipelineExecuteTimeoutError instead, pointing at start_and_wait.

from pipelex_sdk.errors import ApiUnreachableError

try:
    result = await client.start_and_wait(pipe_code="pitch_product", mthds_contents=[bundle])
except ApiUnreachableError as exc:
    print(f"Could not reach the Pipelex API at {exc.api_url} ({exc.code}); check PIPELEX_BASE_URL and your network.")

Public import paths (no barrel)

There is no barrel import — package __init__.py files stay empty. Import each symbol from its module:

• Client & construction — from pipelex_sdk.client import PipelexAPIClient, DEFAULT_API_BASE_URL, MthdsFile
• Run lifecycle types — from pipelex_sdk.runs import RunStatus, RunPublic, RunRead, RunResults, RunResultState, WaitForResultOptions, PollInfo
• Error reports — from pipelex_sdk.error_models import RunErrorReport, UserAction, ProviderErrorMetadata, MigrationErrorBlock, FieldError
• Product wire models — from pipelex_sdk.product_models import UserProfile, MethodData, MethodWriteInput, MethodDraftInput, MethodVersion, MethodVersionSummary, MethodPublished, MethodPublishUnchanged, MethodPublishRefused, Membership, MembershipsResponse, SubscriptionResponse, PlanView, InvoiceView, OnboardingSubmission, UploadInput, UploadedFile, RunHistoryItem, RunDetail, ..., with the catalog-source readers beside them: method_source_to_contents turns a fetched MethodData.mthds into the mthds_contents a run or a validate takes, and MethodFile / parse_method_files / serialize_method_files are the codec for a method's custom PipeFunc python.
• Method selectors — from pipelex_sdk.method_selector import parse_method_selector, ParsedMethodSelector
• Validation verdict types — from pipelex_sdk.validation_models import PipelexValidationResult, PipelexValidationReport, PipelexInvalidReport, ValidationErrorItem, SuggestedFix, VALIDATION_VIEW_INPUT_FORM, ...
• Crate routes — from pipelex_sdk.crate_models import ResolveRequest, CodegenRequest, PipeIORequest, PipeIOValidReport, CrateInvalidReport, MthdsFileItem, ..., the requests and the two 200 arms of resolve, codegen and pipe_io
• Model reference check — from pipelex_sdk.model_reference_models import ModelReferenceVerdict, PresetReferenceVerdict, AliasReferenceVerdict, WaterfallReferenceVerdict, HandleReferenceVerdict, PresetMatch, AliasMatch, WaterfallMatch, HandleMatch, ModelCheckCategory, ModelReferenceKind, ModelReferenceResolution, ..., the verdict of check_model_reference, one arm per reference kind
• Codegen tree — from pipelex_sdk.codegen_writer import write_codegen_tree, CodegenTreeWriteReport to write one, from pipelex_sdk.codegen_check import run_codegen_check, CodegenCheckReport, CodegenDrift, DriftCategory to verify one, with the format primitives in pipelex_sdk.codegen_lock (CodegenLock, parse_lock, load_lock, validate_artifact_path, ...) and pipelex_sdk.codegen_stamp (STAMPABLE_SUFFIXES, is_stampable_artifact_path, compute_content_hash, parse_stamped, ...)
• Typed errors — from pipelex_sdk.errors import ApiResponseError, ApiUnreachableError, PipelineExecuteTimeoutError, PagingNotTerminatingError, RunFailedError, RunTimeoutError, RunLifecycleUnavailableError, RunStillRunningError, CodegenError, CodegenLockError, MethodErrorCode, ...
• Version — from pipelex_sdk.version import __version__
• Client identification — from pipelex_sdk.user_agent import AppInfo, build_user_agent, is_token
• Protocol surface (the MTHDS standard's wire types) comes from the mthds dependency — e.g. from mthds.protocol.exceptions import PipelineRequestError, from mthds.protocol.models import ValidationResult (the neutral verdict union that PipelexValidationResult narrows).
• Input-form descriptors and pipe I/O contracts come from mthds too, because they are the standard's artifacts and this SDK only carries them: from mthds.protocol.input_form import InputForm, InputFormField, ListField, TextField, ... and from mthds.protocol.pipe_io_contracts import PipeIOContracts, PipeInputContract, PresenceMarker, IOMultiplicity, .... PipelexValidationReport.input_form and .pipe_io_contracts are typed with them, as are the three maps of PipeIOValidReport (with mthds.protocol.output_form.OutputForm), so a node narrows on its kind and a slot's presence and multiplicity read as enums — but pipelex_sdk does not re-export the vocabulary, and importing it from here is the one supported path.

Development

make install      # create the venv and install all extras
make agent-check  # fix-imports + format + lint + pyright + mypy
make agent-test   # run the test suite quietly (prints only on failure)
make check        # full gate: agent-check aggregate + unused-imports + pylint

See CLAUDE.md for the coding standards and docs/architecture.md for the design (including the parity map against @pipelex/sdk).

License

MIT — see LICENSE.