Stop babysitting your coding agents.
Talk an idea through once, get back a branch to test drive. Claude Code or Codex builds it unattended.
The runcastle workspace. A sidebar groups six features by who is blocked: two need you, one has an agent working, two are in progress and one has shipped. The open feature is on its tickets phase, with ten dependency-ordered tickets and a next step offering to burn them.
The agent was never the slow part.
A bar showing where the time on a feature goes. Two wide blocks are yours: deciding what to build, and checking what came back. The thin block between them is the agent writing the code.
The orange sliver is the agent writing the code. runcastle takes on the rest.
A project manager that has read everything.
One standing chat sits above your features. Bring it what is bothering you and get back features, in order.
An example. You bring one sentence with three complaints in it. The project session checks the record, finds a shipped feature that left part of the work undone and a feature in review that touches the same area, then cuts the work into one started feature, two parked drafts, and one item sent to an existing feature.
- Knows what every agent did
- Splits the work, sets the order
- Says no to what is already built
- Flags collisions before they merge
The review is a drive, not a diff.
A diff cannot show whether the thing works. So the agents run the app, and then you do.
The review player, paused at 0:42 of the lap 1 walkthrough. The recorded app shows a comment box with a mention typeahead open. A red rectangle is drawn around the typeahead, the drawing tools float above the frame, and a note reading "Typeahead closes when I press the down arrow" is ready to save.
-
An agent drives it first
-
You point at what is wrong
-
The note becomes the next lap
Six phases. You make two calls.
runcastle always knows what the feature needs next. It asks you for two things: approve the plan, approve the merge.
Talk the idea into shape
A real agent terminal opens with the feature already loaded. This part is yours.
The decisions get written down
The same session writes the spec and commits it to your repo.
Skim the task list, say go
The spec becomes small tasks, in the order they have to happen. Click once.
Close the tab
Agents take every unblocked task at once, inside containers. Go start the next feature.
Take it for a drive
The finished branch starts on its own port. Click through the real thing before it lands.
It merges, and it remembers
The branch lands on main. The spec, the decisions, and the run history stay in your repo.
Run more features than you could watch.
Each feature asks for you twice. Six of them is a short queue of decisions, not six terminals to poll.
Six features, each on its own branch and each at a different phase. Realtime team collaboration is on tickets and Entry tags is on review, both waiting on you. Offline draft sync is building with an agent working. Kickoff probe is on spec, Weekly digest email is on ideation, and Mood filter has shipped.
Your agents. Your machine. Your repo.
A small server and a browser tab on localhost. No runcastle account, and no server of ours.
Your agents
Your own login, on the plan you already pay for
Docker
Podman
One container for each unattended build
Your repo
runcastle is built with runcastle.
Counted from its own git history since July 2026.
- 100
- features shipped through this pipeline
- 1,342
- ticket commits written by agents
Get started
One global command. The doctor names anything you are missing.
bun add -g runcastle
Install
runcastle doctor
Check prerequisites
runcastle
Boots on localhost:4512
Needs Bun, Git, and a login for Claude Code or Codex. Full prerequisites