Software intelligence for your agent.
Prepare the change, make it, then review and ship.
Prism is a one-stop software intelligence system that helps your AI agents write better code. It prepares the files that need to be edited. House style helps the agent make the smallest change that matches this repo. Prism then reviews the code, finds bugs, and gives you a ship or no-ship result.
Works with the coding agent you already use.
src/billing/totals.tsEditFollow invoiceTotal in this repo.
Also used in src/checkout/summary.ts.
Receipt: NOT PROVEN
One conversation: the files to open, the edit, and a receipt.
Three steps.
Install
Install the CLI, then run prism install all. It connects every agent on this machine. Open the agent again.
Match the code
While the agent writes, Prism points at a file from your repo to copy. The new code follows that file.
Review
Guardrails, QA, and Architect. The receipt says SHIP, NOT PROVEN, or NO-SHIP.
Prepare, before the edit
Names the files to open, the tests to run, and code already in the repo to reuse.
- src/billing/totals.ts: matches compute, total
3: function computeTotals(lines: Line[])
Tests: test/invoice.test.ts, test/totals.test.ts
Run: npx vitest run test/invoice.test.ts test/totals.test.ts
Match the code, during the edit
Points at a file already in your repo to copy, so the new code looks like the rest. Never a fixture or a test.
src/billing/invoice.ts · invoiceTotal
Not a fixture. Not a test.
Review, after the edit
Guardrails, QA, and Architect. Findings are Critical, High, Low, or Nit. The receipt says whether to ship.
Prism receipt: changed 1 file, 5 dependents, 2 tests reach it. NO-SHIP
Proven 2 scoped test files · typecheck ✓
Failed test ✗ Failed Tests 1
Every caller
Ask where a function is used. Prism lists each caller from its map, so the change reaches all of them.
getUser function · src/users/api.ts
← src/pages/checkout.ts references · measured
← src/users/profile.ts references · measured
Questions Prism answers.
Prism answers from a local map of your repo and runs your own checks. No model is involved.
Prepare hands the agent the few files that matter, the tests to run and code to reuse.
Match the code points at a file from your own code to copy, never a fixture or a test.
Review reads the change and groups findings as Critical, High, Low, and Nit, then gives a Receipt.
Blast radius lists every caller and test a change reaches, from the map of your code.
Guardrails list every edit site, say what still imports a file, and stop edits to generated files.
When the agent stops, Prism runs your typecheck, lint and the tests the change reaches, and the receipt says SHIP, NOT PROVEN or NO-SHIP.
Dev follows the code already in the repo and the blast radius. QA and Architect check the change for bugs and structure.
Write a habit like “deploy” once. Say its name and your agent runs the steps.
Measured on real fixes.
From the Prism 2.0 evaluation on 220 tasks in six public repositories. The method and every result are on the benchmarks page.
See the benchmarksof the files the real fix changed were in Prism’s pack (220 tasks, 6 repos)
of the tests a real fix needed were found by Prism (same 220 tasks)
agents pass install, doctor and a sample hook
Privacy is the default
Runs on your machine. No account, and your code stays here. A usage ping sends the command or tool name and whether it worked. `PRISM_PULSE=0` turns that off.
Works with your agent
Any coding agent that can call tools. One install connects the agent already on your machine.
Open source
Free and MIT. Read the code or build it yourself.
Try it in your repo.
Install Prism, then connect every agent on this machine. Open the agent again.