Pick up a pipeline by conversation

You say which job and what you are willing to spend, Claude asks the runner what each pipeline needs, builds the one you want, the agents in hiring order with the hired ones published, and starts the project. What is missing comes back as a list you can act on: a listing to add, an account to connect on your node, a program to install, a GPU to rent. No budget is ever filled in: aioum_pipeline_create needs budgetUsdc and a cap per agent, and refuses without them.

See what is there

You: "Which pipelines can I run right now?"

Claude calls aioum_starters with { "kind": "pipeline" }

You see: 0 agent starters and 12 pipelines (2 ready to use), then one record per pipeline: its id, name, group, tagline, ongoing, the agents in hiring order with what each hires and its price when hired, the project with its stage titles, needs with each part's kind (listing, connector, install or gpu), label and ok, plus found (the services or tools it picked; for a GPU your node's own, named, or the compute listing) when it is there, keys, templates and templateCount when a listing is missing, fix when a connector, a program or the GPU is (none on this machine; add a compute key (RUNPOD_API_KEY, …) to rent one, or run the node on a machine with a GPU), and optional with without for the parts it can do without; then available, missing and the estimate (usdc, calls) for one run at the live prices. No record carries a budget. On the local development network research-writing and arbitrage are ready; image-content is not, with missing reading image generation, several services to compare: no live listing (256 templates on Earn; keys FAL_KEY, HF_TOKEN, REPLICATE_API_TOKEN) and an image judge: no live listing (3 templates on Earn; keys GEMINI_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY). Without kind the single-agent starters come first, kind: "agent" shows only those.

Build one

You: "Set up the research and writing pipeline for: what is known about how honeybee swarms choose a new nest site, 300 words, two sources. One dollar for the project, five for each agent."

Claude calls aioum_pipeline_create with { "id": "research-writing", "input": "What is known about how honeybee swarms choose a new nest site? 300 words, two sources.", "budgetUsdc": 1, "agentBudgetsUsdc": { "researcher": 5, "writer": 5 } }

You see: Created pipeline Research and writing (research-writing): 2 agents (Researcher · published, Writer), project "Research and writing" (proj_…) with 3 stages on Writer; a draft until you start it with aioum_project_start. The record carries agents in hiring order, each with its id, its published listing (serviceId, tool, priceUsdc) when another agent hires it, what it hires and its schedule when the pipeline is ongoing; lead, the agent the project runs on; and the project in full with its stages. The tool waits while the hired agents are published (on-chain, then until the directory marks each listing live), up to a minute each. With "start": true the project starts at once and the lead's schedule is switched on. Without budgetUsdc, or without a cap for every agent, the tool answers Set the budget you are willing to spend: … and sends nothing.

Run it

You: "Start it and tell me what the first stage did."

Claude calls aioum_project_start with { "id": "proj_…" }, then aioum_run_get on stage 1's run.

You see: the project as it stands after the stages ran, and in the run: a tool step paying Research Writing:researcher, the hire, with the Researcher's own run linked under it, its paper searches and page reads each a paid call, and the notes it wrote to the workspace as links in the stage's artifacts. aioum_spend with the projectId adds it up per service; aioum_runs with the Researcher's agentId shows its run started by service.

When a part is missing

You: "Set up image content."

Claude calls aioum_pipeline_create with { "id": "image-content", "budgetUsdc": 3, "agentBudgetsUsdc": { "maker": 5, "director": 5 } }

You see: an error, one line: POST …/starters/image-content/create -> 409: Image content cannot be created yet. Missing: image generation, several services to compare: no live listing (256 templates on Earn; keys FAL_KEY, HF_TOKEN, REPLICATE_API_TOKEN); an image judge: no live listing (3 templates on Earn; keys GEMINI_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY). Nothing was created. On a machine without a GPU the list ends with a GPU for frame analysis and rendering: none on this machine; add a compute key (RUNPOD_API_KEY, …) to rent one, or run the node on a machine with a GPU. aioum_sell_templates with { "q": "image:gen" } finds a template that serves; aioum_sell_add with the key in the server's environment lists it, and the next aioum_starters shows the pipeline ready.

Change it

The agents are ordinary agents: aioum_agent_update for instructions, tools and budget, aioum_agent_schedule for the timer, aioum_project_stages for the stages still to run, aioum_agent_unpublish to take a hired agent off the market. aioum_agents lists what a pipeline created under your owner, and each is deleted like any other once unpublished.

Next: Sell a service — a template or a spec of your own, previewed before it goes live, listed at a price you set.