Building Phonton: a local-first AI coding CLI that verifies diffs before review

Matthew — Sun, 03 May 2026 04:33:25 +0000

I have been building Phonton CLI, an open-source local-first AI coding agent.

The idea is simple: instead of treating an AI coding tool like a chat box, Phonton treats it like an engineering workflow.

It tries to follow this loop:

You give it a goal.
It creates a plan.
It works against your local repo context.
It verifies the generated diff.
You review the result before accepting it.

Most AI coding tools are optimized for speed and convenience. Phonton is more focused on making AI-generated changes easier to trust.

Why I built it

I wanted a CLI agent that was:

local-first
BYOK, so you use your own provider keys
review-oriented
built around verification
able to remember repo decisions locally
useful from the terminal

The goal is not to replace every tool like Claude Code, Codex, Cursor, or OpenCode. The goal is to explore a stricter workflow for AI-assisted coding: plan, verify, review, then ship.

Current status

Phonton is still early alpha.

It currently has:

a Rust CLI/TUI
provider adapters
planning
local memory
repo indexing
verification flow
review commands
checkpoint/rollback work

There are still rough edges, and I am not claiming it is production-ready yet.

Try it

npm install -g phonton-cli
phonton doctor
phonton plan "add input validation to config loading"

GitHub:

https://github.com/phonton-dev/phonton-cli

I would appreciate feedback from developers who already use AI coding agents and care about local-first workflows, reviewable diffs, and verification.

SIGNAL; shrinking agent skills so context stays for code

Matthew — Sat, 18 Apr 2026 16:49:12 +0000

SIGNAL is a small open-source agent skill bundle for people who hit context limits: tiered modes, symbolic shorthand, checkpoints, a disk-backed .signal_state.md, and separate skills for diff/search summaries and git (commit / push / PR) plus structured review. Each skill ships as readable *.md and a tight *.min.md so you can load less instruction surface when you want.

Install

npx skills add mattbaconz/signal

Repo: github.com/mattbaconz/signal

Benchmarks (honest caveat)
The repo includes scripted fixtures and heuristic token estimates (ceil(chars/4)). That’s useful for comparing shapes, not for quoting exact billed tokens from a provider.

What I’d like from readers
If you’ve tried skills packs or fat AGENTS.md files: does the canonical + minified split match how you’d maintain this, or is it too much friction? I’m open to blunt feedback.