Skip to main content
Plans: Pro and Enterprise. Self-hosted: Included in the open-source server.
On Free and Individual, you can run the sample workflows your workspace ships with, and past runs stay readable. Creating, generating, changing, and deleting workflows, and running your own, is refused with a prompt to upgrade. A routed chat request picks one agent. A MAF (Multi-Agent Flow) workflow is a named, ordered sequence of steps, each bound to an agent, where each step’s output feeds the next and a final step synthesizes one answer. Use a workflow when a task needs several specialists chained — research, then analyze, then summarize.

Define, generate, run

Define — POST /api/maf/workflows with a name, optional description, and steps. Each step needs a task description and either an explicit agent or nothing — unassigned steps are auto-assigned by the routing engine at creation time, scoped to agents you can access. Generate a draft (optional) — POST /api/maf/generate takes a natural-language description, looks at your registered agents, and returns a draft: name, description, output-synthesis guidance, and steps with agents assigned. Review it, then create it for real. The draft is written by a model call on your saved OpenAI key, or the server’s own if you haven’t added one. If no key is available or the provider rejects it, the call returns 502 with code llm_key_rejected. If the model can’t produce a plan, it returns 422. See Generate a workflow plan. Run — POST /api/maf/workflow/{id}/run queues an execution and returns 202 with an execution_id. Poll GET /api/maf/workflow/result/{exec_id} for status and output. An execution is pending, running, awaiting_human, success, failed, or stopped; the last three are final. See Execution status.

The same thing in the dashboard

Everything above is available under Orchestrator → Workflows without writing a request.
Authoring a workflow is the access-control boundary. Which agents a workflow reaches depends on who created it and which agents they named. Treat workflow creation as a meaningful permission, and review agent assignments in any workflow built programmatically.

What a step carries

At run time the platform expands your step’s task description into a prompt — resolving placeholders from earlier steps’ outputs — plus an “extraction goal” describing what to pull from the reply. Each step moves pending → running → success (or failed, or awaiting_human while it waits on a human decision), with token usage and latency recorded. The workflow’s output_generation guidance says how to turn all step outputs into one answer.
Each step adds planning, prompt-filling, and extraction around the agent call, so a workflow execution uses noticeably more tokens than the same number of direct chat messages. Check per-step token numbers before scaling up.

Reliability

Runs are queued and picked up by a background worker, not executed inline. A failed run is retried as a whole execution, up to MAF_MAX_ATTEMPTS attempts in total (default 3). A run that times out, or fails after resuming from a human’s answer, is not retried. Executions in flight during a server restart are picked back up.
A queued execution lives in Redis until a worker claims it. Size Redis durability accordingly — see backup and restore.
Each step lands as its own trace, so the same agent appearing in two steps has distinguishable token usage. See observability.

Example

Two steps — the first names an agent, the second is auto-assigned:
The same calls are available as nasiko maf CLI commands — see workflow CLI reference.

Routes

Routing and flow limits

The routing engine that auto-assigns unassigned steps.

Flows in the dashboard

Watch a multi-agent run step by step.

Observability

How per-step traces stay distinguishable.

ADLC

Building the agents a workflow calls.