Actions and REST calls
Describe an HTTPS call without code and Weave turns it into a reviewable Action on its built-in HTTP connector. OpenAPI documents can supply the operations.
Read the guide
Business processes that cross systems and people.
Draw a workflow in Studio or write it in YAML, JSON, or Python. Its runs wait for people and events while Weave records their history step by step.
How it works
How a workflow becomes a run you can follow.
YAML, JSON, and Python versions of a workflow compile to the same artifact digest before anything runs. Editing a draft never changes a run in progress.
Author
Write the input schema, the steps, and the output schema, in Studio or in a file.
Compile
The compiler checks schemas, Actions, and values against a catalog, then produces an artifact with a digest.
Publish
Publishing stores an immutable version in a project. A change means a new version.
Activate
Activation pins a version to an environment with its connections, worker releases, and human-task assignments.
Run
An input sent to an activation starts a run with its own state and history.
Observe
Read a run’s status, output, history, and incidents in Studio, from the CLI, or through the API.
Building blocks
The steps a workflow is built from.
A workflow is a tree of steps: a Decision or a Parallel group always joins again before the next step. Values come from expressions, since Weave has no scripting language.
Workers
Your integration code runs in workers, processes that claim a task, call your system, and report the result through the API, with no database access.
Read the guideHuman tasks
A human task waits for one authenticated person to choose an allowed decision on a form. Each environment binds the task to people or groups.
Read the guideEmail conversations
Weave sends mail through SMTP and reads a mailbox over IMAP without changing it. A reply carrying a valid routing token signals its run.
Read the guideDecision tables
A decision table is a versioned policy with typed input and output that Weave evaluates inside the run, with no worker or provider call.
Read the referenceEvents and schedules
Schedules start runs on a UTC timetable. Signed webhooks, Kafka records, and email or messaging sources start a run or signal one.
Read about eventsFiles
Steps pass verified file references. Weave stores the bytes apart from history, which stays readable. Replay never transfers a document again.
Read the guideAI tasks
An AI task’s answer counts only when it matches its author’s contract. A Firefly Agentic worker makes the model call.
Read the guideLoops and sub-workflows
A Loop over items step repeats its steps for each list item. A Call a workflow step runs a published version as a sub-workflow.
Read the step contractsDurability
A run keeps its place through restarts and long waits.
The Weave API stores workflow versions, run state, and tasks durably. Waits, signals, and human tasks are stored with the run and survive an API restart.
History records what Weave accepted for a run, step by step. Recorded replay reads it offline without contacting anyone.
Weave delivers external calls at least once, because a worker can crash after another system accepted its request. Steps that change another system must be idempotent. Weave retries an Action only when its declared side effect makes a repeat safe. Otherwise the run is suspended with an incident until an operator decides.
Read about tasks, leases, and retriesCode
A customer check that waits for an approval signal.
The repository’s customer-onboarding workflow asks another system about a customer, then waits up to a day for a manager’s decision.
Check the customer
checkcalls a published Action that asks your system whether the customer is eligible.Wait for approval
approvalwaits for thecustomer-approvedsignal, whose payload must match its schema.Decide
decisionis a Transform that accepts the customer only when both answers are true.
In the Python SDK, WorkflowBuilder produces definitions like this one as data. WeaveClient publishes them and starts runs through the HTTP API.
customer-onboarding.workflow.yaml
steps:
- id: check
kind: action
uses: onboarding.check-customer@1.0.0
with:
object:
customerId: {ref: /input/customerId}
- id: approval
kind: signal
name: customer-approved
timeoutSeconds: 86400
payloadSchema:
type: object
additionalProperties: false
required: [approved]
properties:
approved: {type: boolean}
- id: decision
kind: transform
value:
object:
accepted:
op:
name: and
args:
- {ref: /steps/check/output/eligible}
- {ref: /steps/approval/output/approved}
output: {ref: /steps/decision/output}Deployment
Run Weave as a service or inside your product.
Your product calls Weave under its own identity and grants, and never needs the orchestration database.
Standalone or embedded
Run Weave as its own service, or embed its compiler and workflow builder, which need no server, database, or network.
Read the embedding guidePython SDK and HTTP API
The Python SDK describes workflows, controls a platform over its HTTP API, and runs your code in workers. Other languages call the HTTP API.
Read the SDK referenceYour identity provider
Bring a compatible OIDC or CIAM provider. Weave verifies the provider’s tokens, then checks its own scoped grants for what a caller may do.
Read the identity guideContainer deployments
Operations registers an existing Compose project, Kubernetes namespace, or Azure Container Apps environment. A runner beside it applies the plans people approve.
Read the deployment guideOpenTelemetry
An API process can export metrics and traces to an OpenTelemetry collector over OTLP. Export stays off until you set WEAVE_TELEMETRY.
Studio’s Operate area
Operators cancel a run, retry a finished one, or resolve an incident from Operate’s lists of runs and incidents.
Read the incident guideFoundations
Built on pyfly, with AI work on Firefly Agentic.
Weave is a native pyfly application: dependency injection, controllers, and request filters come from the framework. You can embed the compiler on its own: importing it doesn’t start the application.
An independent Firefly Agentic worker runs AI tasks. It calls the provider and model that the workflow’s AI profile specifies.
The Firefly Software Foundation publishes Firefly Weave under the Apache License 2.0.
Sources and attribution
Copyright 2026 Firefly Software Foundation. The workflow excerpt is shown unmodified.
Next step
Simulate a workflow on your computer.
The Weave documentation covers installing the CLI and writing the workflow, and simulation needs no server.