Runkite

Open agent control plane

Live Admin at getrunkite.com/admin. A real Runkite process, not a mock

live plane · same binary as runkite serve Open Admin →
Runkite Admin walkthrough: overview, agents, runs, connectors, grants, HITL, kill, audit. Click to open the live plane.

Scenario

Three agents. Two frameworks. One connector.

When one script stops being enough: coordinators and workers still need one place for credentials, HITL, kill, and audit.

Without a plane

Three scripts, three copies of the secret

Each process owns its env, HITL, and logs. Delegation is ad-hoc HTTP. Ops is SSH and grep.

With Runkite

One plane. Run-bound sessions. Shared Admin.

Clients speak Agent Protocol. Runners pull jobs. Connectors mint credentials from the assignment — not from forged headers.

Try it

This Admin is the plane

Click the walkthrough or Open live Admin. You are hitting a running Runkite binary on this host (Postgres, Redis, runners), not a screenshot. Shared key below. Fake agents, no SLA. Your real graphs stay on your runners: Compose, Helm, Kubernetes, a VM.

600d9704de73d08db3b5d621b6e17c9172053cb480415ae8

Open live Admin Install (Compose / Helm)

On the plane

Governance that fails closed

Grants, HITL, fencing, and reclaim ceilings live on the plane — not as afterthought middleware in every script.

Connectors Minted per run. Workers never hold long-lived provider secrets.
HITL Approve or deny before a marked connector call proceeds.
Fencing A superseded runner cannot overwrite the winner after reclaim.
Poison pill Crash-loop jobs stop reclaiming after a hard ceiling.

Architecture

Clients never talk to runners

Agent Protocol on the front. Runner Protocol on the back. The plane owns lifecycle, policy, and proof. Pluggable state, transport, and vectors under an honest tier model — Supported for Postgres + Redis; Compatible for the rest.

Start

See Admin, then deploy the plane

Sandbox here. Supported production is Postgres, Redis, and a runner you run.