TarkOS Cockpit

The operations console for my agents: what ran, what needs me, what to decide.

A desktop console over TarkOS's files. It answers four questions: what ran while I was away, what needs my decision, what the agents are doing now, and where each project is heading.

01 · Views
23
views over the harness, from one route config
02 · Backend
108
backend commands between the UI and the files
03 · System map
270
harness components drawn in the system map
01 · Problem

Hundreds of changes overnight. A handful of real decisions, buried.

When the agent works a backlog while I’m away, the morning brings a long log. Most of it is reversible and fine. A few items genuinely need me, and in a plain log they look exactly like everything else.

02 · What it does

Receipts for what’s done, cards for what needs me

  1. C1: Receipts

    Reversible, high-confidence actions apply themselves and leave a receipt I can revert.

  2. C2: Decision cards

    Irreversible or taste calls wait as cards: the options, the agent’s recommendation, its stated confidence, and one-click ratify.

  3. C3: Calibration

    A separate view checks whether the confidence is honest: does a stated confidence match how often the agent turns out to be right? That decides how much the agent may do on its own.

  4. C4: Sessions, projects and the system map

    Live sessions with their context use, a goal tree per project, and a generated map of every mechanism, skill, hook and tool in the harness.

03 · How it's built

A thin app over git-backed files

There is no database. The harness keeps its tasks, decisions and project state as files in git, and Cockpit reads and writes those same files: a Rust backend with 108 commands, a file watcher that pushes changes to the UI, and small command-line tools that turn harness data into JSON. Ratifying or reverting a decision writes back to the file the agent reads next, so the console and the agent can never disagree about state.

The code is written by AI agents under my specification, test contracts and review of the result, like the rest of the tools around TarkOS.

04 · Key numbers

Measured, with dates

Views
23

Views over the harness, all from one route config.

Backend commands
108

Commands between the UI and the harness files.

System map
270

Harness components drawn in the generated map.

History
1,050+

Commits since April 2026.

05 · Screens

Real captures, synthetic data

fig.1 · cockpit · decisions real capture · synthetic data
TarkOS Cockpit's Decisions view on synthetic demo data: decision cards with options, a recommended option with a stated confidence and a Ratify button, counters for decisions waiting and auto-applied actions, and an autonomy ladder summary. Click to zoom
FIG. 1 · Decisions that wait for me, with the agent’s recommendation and confidence. Try the live demo (opens in a new tab) (synthetic data).
fig.2 · cockpit · today real capture · synthetic data
TarkOS Cockpit's Today view on synthetic demo data: today's receipts (sessions closed, commits, decisions closed) with a timeline of autonomous runs, a day plan, and a Needs you column with decision cards showing the recommended option, its confidence and a Ratify button. Click to zoom
FIG. 2 · The day view: receipts from the autonomous runs, and the decisions waiting for me.
fig.3 · cockpit · calibration real capture · synthetic data
TarkOS Cockpit's Calibration view on synthetic demo data: prediction hit rate by session, observed accuracy per stated-confidence band, per-category accuracy that feeds the autonomy levels, and the authority map from A1 to A4. Click to zoom
FIG. 3 · Calibration: stated confidence against observed accuracy, per category.
06 · Status and next

Maintained, not the current focus

Cockpit was built out in June and July 2026 and is maintained since; the harness itself is where my time goes now. The captures above and the live demo use a fictional company’s data, so nothing from my own projects is shown.

Talk to me about this →