Rules are optional. Checks aren't.

For everyone who has typed “ALWAYS run the tests” in capitals. Trinket turns rules into checks that run in your agent's hooks.

Preview build · Apple silicon · macOS 27 or later · Bring your own AI account

AGENTS.md
  1. YouRun the tests before you say done. Agent✓ all tests pass! Trinketno test ran. back to work.
  2. YouALWAYS commit with the git skill. Agentskipped it, was faster Trinketblocked. use the skill.
  3. YouSubagents should follow these rules too. Agent✓ passed it on. Trinketcaught one skipping it.

Your agent has excuses. Trinket has receipts.

“All tests pass.” You never ran a single one.

Agents grade their own work. Trinket doesn't.

When your agent says it's done, a check reads the evidence first. It looks at the commands that ran and the files that changed. A separate model gives the verdict.

turn-1.2.0.trinket293 KB

Did that plugin just read my API keys?!

Plugins see your keys. Workflows don't.

Your AI keys stay encrypted on your Mac, locked by your Keychain. Workflow code runs in a sealed Linux box that reaches only the sites it names.

You ran for an hour. Doing WHAT?

Chat logs scroll away. Run records don't.

Each run keeps its plan, every step and every AI call. Look back at any run, even after an update or an uninstall.

It updated itself overnight. Now nothing works.

Plugins change overnight. Pinned workflows don't.

Published workflows arrive signed. Let publishers you trust update automatically, or pin a version until you're ready.

Half my context is rules. You ignore every one.

Rules take up context. Checks don't.

A check runs outside your agent's conversation. Only its verdict comes back, so your context stays free for the work.

Switched to Codex. Rebuilding everything. Again.

Hooks are per agent. Workflows aren't.

Every workflow is an MCP tool and a terminal command. Switch from Claude Code to Codex, and the same checks still run. Only the hook wiring changes.

A receipt: one finished run

Research Answer Overview Runs
Runs
Research Answer Done · 45 s
Done what it was asked
How does MCP use PKCE to protect authorization codes?
Path each step it took
AICollect evidence CodeJudge evidence
Fill gaps · skipped
AIWrite cited report CodeRender citations CodeScore report
Time per step
  1. 1research36 s
  2. 2judge1 s
  3. 3report6 s
  4. 4render0.5 s
  5. 5score0.9 s
✓ Result 3 outputs
Report what came back How MCP uses PKCE PKCE ties each authorization code to a secret the client made. A stolen code is useless without it. [1]

How it works

Install a workflow once. Then start it from the tools you already use.

  1. 01

    Install a workflow

    Double-click a .trinket file. Trinket checks the package, and its publisher's signature when it has one, before anything runs.

  2. 02

    Start it from anywhere

    Ask Claude or Codex to run it, or start it from Terminal. It keeps going even if you close the chat.

  3. 03

    Read the receipts

    Open the run to see each step, what the AI was asked and what came back. Results stay readable even after you remove the workflow.

Start with your own workflow

Tell Claude or Codex what you want automated. It reads Trinket's built-in guide and writes the workflow for you.

Trinket runs it sealed, so you can try code you didn't write. When it works, share it as a single .trinket file.

Terminal
# one-time setup for authors
$ trinket download-tools

# start, install and run a workflow
$ trinket init my-flow
$ trinket install my-flow
$ echo hello | trinket run my-flow.echo --wait

# package it to share
$ trinket build my-flow

What you need

  • A Mac with Apple silicon
  • macOS 27 or later
  • Your own AI account, such as OpenRouter or a ChatGPT plan

Does my data leave my Mac?

Only to the AI account you connect, and to sites a workflow declares. Trinket has no cloud service and no sign-in of its own.

Can a workflow read my files?

No. Each step sees only the text you hand it. Writing a result back to a file is your call, and Trinket won't overwrite a file that changed in the meantime.

Can my agent get around a check?

Sometimes. Checks run through your agent's hooks: in Claude Code, and for the done check in Codex too. A command hidden inside bash -c or an alias can slip past the commit check. Some checks only report. A check that can't run lets the agent through.

Do I need to code?

Not to install and run workflows. Making one takes an AI coding assistant, such as Claude or Codex, or some TypeScript.

What does it cost?

The app is free. Workflows are what you'll pay for, and the official ones are free during the preview. You pay your AI provider for what the models use.

Hand Trinket your next task.

Install it, add your AI account, and describe your first workflow.

Download for Mac

Preview build · Apple silicon · macOS 27 or later