Jayphen

Today I am working on a benchmarking dashboard to benchmark Coducky code review performance against other major competitors. I am using Martian's code review benchmark: github.com/withmartian/code-review-benchmark

Unlike the competitors, in Coducky you can choose your own models. For a quick check of the benchmark against 5 PRs I used Qwen 3.7-plus in single-panel mode (as opposed to multi-panel, where you can choose many models to review). F1 is the important metric here as the harmonic mean of precision and recall.

With this cheap model (about 2c per PR review), Coducky's harness is already scoring higher than most of the major players.

I am running a full 50-PR benchmark now (ETA 2.5 hours) and will see how we fare across the entire corpus.

Most AI code review platforms are off-device, run on PRs and integrate with GitHub (e.g. Coderabbit). With Coducky I'm trying to do the opposite and focus on fast, cheap, on-device iteration. Part of what I am thinking about now is how to leverage on-device models as part of the local code analysis layer. This reduces costs for the end user and can also enable precomputation while idle.

I'm looking at integrating Apple Intelligence alongside deterministic analysis using tree-sitter and graph/vector retrieval. Today I'm working on a spike to de-risk some of the ideas I am thinking about. I'm going to first try to use Apple's Foundational Models to generate branch descriptions automatically in the background based on branch/ref labels, commit subjects, diff stats, changed paths, etc.

I said I'd ship it today and I did: MCP support is done! Users will be able to drive a code review from any MCP-enabled client.

I also put together a doc page to explain how to connect at coducky.com/docs/mcp-clients/

If you look carefully, you may just find a download link to the pre-release app.

By the end of today I will ship MCP support with a new docs page that explains how to use it in all clients.

coducky.com join the list!

Today I am working on some UI and workflow improvements.

I've added a bottom breadcrumb bar and a way to navigate between commits easily.

I also made the AI Review button a primary action.

I plan to also redo the titlebar and sidebar to follow Apple's HIG.

I shipped the landing page today. Go and sign up for the wait list to lock in the launch pricing! coducky.com/

I am just waiting on Lemon Squeezy to approve my store so that licenses can be purchased, and then I will make the app public.

I also have added a public changelog, which you can find on the website.

The website (Astro) is in a separate repo to the app (Swift) as I didn't want to manage a monorepo. I maintain a changeset in the main repo and when I tag a new release I upload it to R2 during CI and trigger a rebuild of the Astro website, which pulls the changelog during build to create the changelog page you see above.

Today I got Sparkle working, released my first patch version, added model selection autocomplete, added Cursor and GitHub Copilot support, and refreshed the visual design. I also have a draft of the website ready.

Starting to work on the landing page for Coducky today. I'm going for a native macOS aesthetic because I want to signal that the app is native. My intuition is that native apps are valued more highly today because people are producing more cross-platform apps than ever with LLMs.

I bought a Mac Mini the other day (soon will be handing my work laptop back as I was made redundant 🥲) and have just been setting it up as a self-hosted runner for GitHub Actions. It is working great already!

- No more exorbitant GHA fees due to MacOS runners being 10x Linux cost

- It's way faster to run on my machine than Github's

As a result I have my first 0.1.0 release of Coducky. Tomorrow I will get Sparkle working, clean up some UI elements to make onboarding a little more clear, and then I will start distributing trial versions!

Today I've been working on implementing tool use / reasoning output as agents are reviewing the code. I previously had hidden this and just shown a loading spinner, but I feel exposing it gives more confidence that the agents are thoroughly reviewing.

My goal is to make the app available in the next week at a lifetime $49 license, with a 7 day trial. I have some admin to do first, and I still need to build a landing page, but I think I can get there.

Today I went to the pub and also managed to get the core of the ACP implementation working. This means you can combine your subscriptions (CC, Gemini, Codex) with OpenAI-shaped providers (OpenRouter etc).

The goal here is local reviews of your code with as many agents as you can afford. Subscription subsidisation make reviews a lot more affordable!

Running sequential code reviews with different LLMs has been proving its worth over the last couple of days since I implemented the feature. I am finding that follow-up reviews are correctly challenging the initial runs, and producing a better result overall.

Today I have been working on adding tool calling for agents so that they can grep, ls, and readfile outside the touched files.

One of the tradeoffs I am trying to balance is token usage. I am seeing reviews range from $0.03 - $0.25 depending on the size and model (using grok-4.3 and qwen-3.7-max via openrouter). I may need to introduce levers so that users can choose whether to do deeper reviews with full tool usage and post-change file contents, or more shallow reviews that are diff-only.

I'm also looking in to adding support for ACP agentclientprotocol.com so that subscriptions (Claude Code, Codex, etc) can be used. I already have MCP support, but ACP would allow the user to kick off a review from inside the app (MCP is the opposite direction)

Today I've been working on sequential LLM reviews for Coducky - so the LLMs can respond to the feedback that others give.

I need to spend a little more time on the UI here - the outline sidebar I think may be wasted space.

Dogfooding my second app as I prepare it for launch as well. This one will be a lifetime license. It's for native AI-assisted local code review - bring your own key or use the MCP to generate reviews.

RevenueCat is wired up.

I just need to create an onboarding flow and paywall screen design and I think I might be ready to ship macOS, Windows, and iOS. For Android I believe I will need 12 testers before I can ship there