discern v1.0.0

Intelligence, in practice.

discern installs a disciplined engineering practice into your software project. You bring the vision, your agent writes the code, and your project gets better at being built.

Get started See how it works

Any stack. Runs offline. No API key.

Works with the coding agents you already use

Claude Code
Codex
Gemini
Cursor
GitHub Copilot

Set the direction.

Keep your decisions and expectations in the project, where your coding agents can put them to work.

Let more work move.

Give each task its own space, so several agents can move different parts of your project forward.

Know what’s ready.

Review work with evidence from your project's own checks. You decide what ships.

The practice

Two kinds of intelligence. One shared project.

An engineering practice for agent-built software, installed in your project. You set the direction, your agent carries the work, and the project keeps what matters between you.

You decide what quality means for your project. discern keeps that judgment in the project itself, where every coding agent you use can find it and put it to work.

Your agent operates discern day to day. It starts each task in its own workspace, runs the checks your project declares, and brings the work back with evidence that they passed. You review the result and decide what ships.

For people who take their software seriously.

If you have only ever built software through an agent, discern asks nothing of you beyond the decisions that are yours to make. Your agent sets it up and runs it, and the discipline is in place by the time people start depending on your work.

If you already direct more implementation than you can personally read, discern lets your judgment reach every change without you reading every line. Run more work in parallel, keep your standards across models and providers, and stay the one who decides.

Built around the way agents work.

Most developer tools assume a person at the keyboard. discern treats your coding agent as its day-to-day operator, and most of what it does for the agent happens out of your sight.

Every session opens already briefed by your project's compiled instructions. One call reports what is true right now and the next valid step, so the agent never rebuilds the picture from a long transcript.

Results come back bounded, so the agent spends its context on your change rather than on the tool. A failed check returns the evidence and a command to reproduce it, and a refusal always names a way forward. The agent operates; you still decide what lands.

Give each task its own workspace.

Every piece of work begins in a separate checkout on its own branch. Several agents can move at once without treading on each other, and nothing unfinished reaches your shared branch.

Know what passed.

Finished work comes back with the results of your project's own checks, tied to the exact change that passed them.

Keep every gain.

When a quality measure improves, the project keeps the new level. A later change cannot lower it.

You decide what ships.

Passing checks makes a change ready for a decision. The decision to land it stays with you.

Your agent sets it up.

Tell your coding agent to set up discern. It studies your project, wires up your real checks, asks only for the decisions that are yours, and proves the setup works before it finishes.