tony

Your AI wrote the code. tony explains what now exists.

tony turns a git diff into a review page for the person who now owns the code — often someone who didn't type a line of it. Not a linter, not a code-review gate: it builds your mental model of the change, fast.

tony does not call a model. It runs inside the coding agent you already pay for: you ask the agent to review a branch, it calls tony over MCP, tony hands it the diff and the standard the review has to meet, checks the answer against the diff, and publishes the page.

Every review page has three tabs:

1. File changes — the diff per file, with plain-language annotations sitting inline above the lines they explain. Modified code gets Prev / New / Changes panes; line numbers are computed from the diff, never trusted to the model.
2. Blast radius — files the change reaches that are not in the diff, each consumer classified breaks / behavior-change / compatible, shown at the affected lines.
3. How it works — steppable runtime walkthroughs. One concrete scenario ("you run export --resume after a crash") traced step by step, showing the real source lines read from your disk, a small state table, and one plain sentence per step.

Install

curl -fsSL https://tony-cli.com/install.sh | sh

or, if you already use uv or pipx:

uv tool install tony-cli     # or: pipx install tony-cli

Either way you get a tony command on your PATH.

Set up

Two commands, once per machine:

tony connect                 # register tony with Claude Code, Codex, Cursor, Amp
tony login                   # sign in, so reviews have somewhere to publish

Then restart your agent and ask it, in words:

review this branch with tony

The agent calls tony_start, reads the diff and whatever files it needs, writes the review, and calls tony_publish. What comes back is a URL. Reviews are published to your account; the page is the deliverable, so the agent hands you the link rather than pasting the review into the chat.

tony reviews committed work: the page shows code straight from disk, so the working tree has to match the revision under review.

Commands

tony --help prints this list.

Every command asks PyPI whether there is a newer release and says so on the way out; the request runs alongside the command, so nothing waits on it, and a review started through the MCP server tells the agent instead of a terminal nobody is watching. It never installs anything by itself.

| tony uninstall | Delete tony and everything it wrote — see Uninstall. | | tony --version | The installed version. | | tony --help | This list. |

tony uninstall takes --yes (-y) to skip the confirmation and --no-scan to limit it to the current repo instead of searching your home directory.

Both tony update and the installer resolve from PyPI explicitly, so a machine pointed at an internal mirror can't decide what tony-cli is.

Uninstall

tony uninstall

Removing the package by itself would leave the parts worth removing: your login token in ~/.tony, and — from versions before tony went agent-native — an API key beside it and a .tony/ directory of past reviews inside every repository you ran tony in, which hold source code. tony uninstall lists everything it found, waits for you to type yes, deletes it, then removes the package with whichever of uv, pipx, or pip installed it.

It searches your home directory for stray review directories; --no-scan limits it to the current repository, and --yes skips the prompt.

Development

pip install -e ".[dev]"      # the CLI, with pytest
python -m pytest             # the deterministic layer's tests
cd web && npm install
npm run dev                  # the site; /dev renders src/fixtures/review.json

The renderer is web/src/renderer/render.ts, and the site is the only thing that runs it. The Python side decides all layout facts (line numbers, spans, added/changed tags) in src/tony_cli/layout.py and ships them in the payload; the renderer never re-derives them. DESIGN.md is the visual contract.

The site deploys to Vercel and needs DATABASE_URL (Neon Postgres), BLOB_READ_WRITE_TOKEN (Vercel Blob), and GITHUB_CLIENT_ID (an OAuth app with device flow enabled).