OpenClaw

The Pump Up plugin lets a self-hosted OpenClaw agent pause on an approval or elicitation, durably park while a human decides, and resume in its live session when the decision lands — without you writing a poll loop. It gives the agent a set of Pump Up tools and runs the wait-and-resume machinery for you.

The plugin is in preview — available as source in the public repo. Packaged distribution (ClawHub / npm) is coming.

What the agent gets

Once enabled, the agent can call these tools:

ToolWhat it does
pumpup_request_approvalAsk a human to sign off; the agent pauses and resumes on the decision.
pumpup_request_elicitationAsk a human for structured input; resume with their answers.
pumpup_record_actionRecord an agent action on the timeline (fire-and-forget).
pumpup_report_exceptionFlag an execution exception.
pumpup_add_noteLeave a note on the timeline.
pumpup_get_decisionRead a decision inline, without parking.

The plugin handles task targeting itself — one durable task per session — so the agent never manages task ids. Attachments are workspace-relative file paths, confined to the workspace directory.

Configuring

Configure the plugin under plugins.entries.pumpup.config:

KeyDescription
baseUrlYour Pump Up API base URL.
apiKeyAPI key (or set PUMPUP_API_KEY).
ownerSessionKeyThe session that owns the outbound poll loop.
pollIntervalMsHow often to check Pump Up for decisions.
maxConcurrentResumesCap on concurrent agent resumes.

Exposing the tools

Plugin tools are gated twice — both are required:

  1. The plugin manifest declares contracts.tools (it does), so the runtime will register them.
  2. The operator allowlists them — add them to tools.alsoAllow, or run the full profile. The default coding profile hides plugin tools, so without this the agent never sees them.

Operator CLI

Inspect parked requests locally, without a backend call:

$openclaw pumpup list
$openclaw pumpup get <flowId>