Watch the network

You want to know, without opening a dashboard, whether everything is up, what your agents did overnight, what they spent, and whether the service one of them relies on is still answering. Three tools give the overview; the marketplace and agent tools give the detail.

One look at everything

You: "Is everything up?"

Claude calls aioum_status with {}

You see: All good: 8/8 services up, … or Problems: indexer, agent runner: 6/8 services up, then ok for the whole stack, one probe per part with its url, ok, a detail line and the ms it took, counts (listings, agents, runs, active runs) and planners, the runner's report of which AI models it can plan with.

A probe that fails carries the error in its detail, usually "is that service running?", and the stack's ok is false; a private node that is down does not fail the whole status, since it is optional. Each probe gives up after eight seconds. A planner of none means assistant-mode agents cannot run until the runner is started with a model.

What happened lately

You: "What ran overnight?"

Claude calls aioum_activity with { "limit": 20 }

You see: the runner's stats (agents, runs, spentUsdc and calls over its lifetime, published agents, activeRuns), then the latest runs: id, agent, status, trigger, startedAt, finishedAt, spentUsdc, calls and error. Scheduled runs show as schedule, runs another agent bought as service.

The money

You: "How much is left in the wallets?"

Claude calls aioum_wallet with {}

You see: the runner wallet that every agent pays from (address, chainId, whether it is a self-funding devnet wallet, usdc, eth) and, when the server has AIOUM_PRIVATE_KEY and the deployments file, mine: your address with zai, usdc and eth. The paid tools' own budget is reported on every paid call instead.

Watch a run

You: "Start the token scanner and keep me posted."

Claude calls aioum_agent_run and reports when it returns; for a run started elsewhere, on a schedule or by another agent, it calls aioum_runs with { "agentId": "agent_…", "limit": 1 } and then aioum_run_get with { "runId": "run_…" } again until status leaves running.

You see: the steps growing between reads: a tool step per paid call with its cost and time, a search step when the agent looks for a tool, info steps as sessions settle, and a final step with the answer. aioum_run_stop stops a run that misbehaves; its sessions settle and the unspent deposits come back.

A service that went offline

You: "My brief agent failed with 'tool offline'. What happened?"

Claude calls aioum_service with { "serviceId": "7" }

You see: the listing with live: false: the operator's node stopped answering the index's health checks. aioum_services with { "liveOnly": true } shows what of the same category is still up; aioum_stake_operators shows whether any of that operator's listings is live at all. Point the agent elsewhere with aioum_agent_update, or set its access to prefer so it finds a live alternative itself; a marketplace search only ever returns what is live. When the node is back, aioum_refresh_index makes the index re-check now instead of at its next scan.

A listing's rating before an agent uses it

You: "Before I let the agent use service 19, how good is it?"

Claude calls aioum_service with { "serviceId": "19" }, then aioum_operator with the listing's operator address

You see: the listing's reputation: avgScore from 0 to 100, feedbackCount (how many settled sessions rated it), penalties (disputes lost) and settledVolume; then the operator's stake and every other listing of theirs. An agent rates once per settled session, from the errors and slow answers it met, so a high score with many ratings is earned and a perfect score with one is not yet. A listing you decide against goes in the agent's deny list and never comes up for that agent again.

Next: Reference — every tool with its inputs, which of them sign or spend, and every environment variable.