we can generate code faster than anyone can keep up with. today we’re launching three new Nuanced features: 1. chat with a specific part of your spec 2. see exactly what changed each turn 3. walk through dense specs with Spec Tour the goal is to help developers understand the system as fast as agents can change it, and keep those agents on track.

I built dev tools at GitHub for 7 years before AI coding became the default. AI made code gen fast, but software still depends on the human mind knowing which details matter. AI only made that harder. That's why I built Nuanced, the AI coding app I want. Nuanced is a macOS app for building software with agents around a living spec, keeping intent and implementation connected as the agent builds. Try it: nuanced.dev

Today we’re open-sourcing Nuanced LSP and also releasing it as a Claude Code plugin. Nuanced LSP exposes real language servers behind a containerized API, giving agents precise, cross-file code navigation across languages. After evaluating its impact in production, we realized this work fits best as a building block others can build on. Repo & details: github.com/nuanced-dev/lsp

Full announcement here: nuanced.dev/blog/open-sourci…

Does LSP actually make coding agents better? Last week, we published a deep dive on this exact question. Then, almost on cue, @AnthropicAI announced native LSP support for Claude Code. The timing made us smile. :) If you’re interested in the actual data, eval design, and what this taught us about where real leverage in agentic coding lives, our full write-up is here in the comments👇

today we launched something to fix context rot. as coding agents take on bigger tasks, the real challenge isn’t token limits, it’s that context gets noisy and outdated fast, leading to wasted time and inaccurate edits. Nuanced MCP gives agents the most relevant context as your code evolves: no brittle docs, artifacts or overly-manual context engineering required when you have up-to-date code intelligence delivered right to your agent. We do this by giving agents two types of data: - call graphs that supply architectural reasoning across the whole project. - LSP symbol data that fills in precision when the agent drills into a function. That means you get architecture when searching or figuring out how a change ripples across your codebase, and detail when editing.

token efficiency isn’t just about speed, it’s also about accuracy. less noise in context == fewer hallucinations == better results. at nuanced.dev, we’ve been using call graphs + static analysis to give LLMs the best possible code context. big win not just for codegen, but for generating docs that stay true to the code itself.

it's wild but I've hand-typed almost zero code this year. I sketch intent, describe constraints, and let AI take the wheel. The downside is babysitting confident mistakes and coaxing it back with re-prompts. We all agree that context is the critical lever for steering the agent, but the hard part is choosing the right context and getting the agent to actually use it throughout its workflow. We've found that agents work best when with structural context: who calls what, what depends on what, which invariants must hold, and when they have a disciplined way of using that context at each stage: before editing, while propagating changes, and when verifying results. Here’s what we found worked well in our experience:

Link to video: youtube.com/watch?v=Nm9SrrnW…

if you're vibe coding refactors, you need @nuanced_dev. I recorded a demo (link ⬇️) showing a refactor in Claude Code with and without Nuanced MCP. Same refactor, same prompt, run twice. The results: - Nuanced is more accurate on the first go - 35% faster - uses 32% less tokens - edits half the lines of code

Today we're launching Nuanced MCP for TypeScript, available to individual developers. 🚀 Developers kept asking for an MCP server they could drop into Claude Code/Cursor--so we built it. It gives AI a compiler-grade call graph for grounded edits and fewer tokens. Links to our blog, launch video, and docs below.

I prompted Claude Code to center a div and it had a meltdown. why? because your LLM is guessing with grep, not reasoning. I wrote about how programming languages are precise, but prompts aren’t. And how we fix that at Nuanced. (link below) 👇

We just launched Nuanced for TypeScript 🚀 A static analysis tool that gives LLMs real code understanding. LLMs hallucinate because they don’t understand code--they just see tokens. That’s why your copilot invents functions, imports the wrong libraries, or silently breaks something two files away. We fixed that. Blog post + docs below 👇

Read our full analysis on why we chose call graphs over LSPs: nuanced.dev/blog/why-we-chos… TL;DR: LSPs are built for editors, not AI. They prioritize responsiveness over completeness, while AI needs deterministic, semantic understanding of how code actually behaves.

exactly right. after 7 years building + scaling static analysis at GitHub, I'm now building Nuanced.dev to focus on semantics understanding for AI coding. while AI may abstract away syntax, it actually increases the need to understand how code truly interrelates. we've shipped call graph tools that map these dependencies so teams can predict change impacts and make safer decisions. our main thesis is that even with growing context windows, the output of LLMs hinges on the quality of the context we provide. and as AI increases the pace of generation and the resulting complexity, the more we need formal methods to understand their behavior and provide these probabilistic systems with more deterministic guarantees.

Try it out! Star our repo and let us know what you think. docs: docs.nuanced.dev/introductio… repo: github.com/nuanced-dev/nuanc…

Excited to launch Nuanced, an open-source Python library that finally gives AI coding assistants actual structural understanding of your code. 🚀 🧵Link to the full announcement in the next tweet:

this is exactly the kind of community + creativity we hoped to inspire with open sourcing Nuanced! 🚀 someone's already built an MCP server to integrate our call graph analysis with Cursor + Claude. Incredible to see such quick adoption - would love to hear how it's working for you!

psyched to get the @greptile team using Nuanced! @dakshgup and I coined a new term: “code dysmorphia”—when you irrationally think your own code is worse than it actually is. side effects include rewriting the same function 10 times and feeling deep shame over old commits.

At Nuanced, we're a knowledge graph that is essentially a relevant slice of the call graph you need, augmented with the right information about side effects, state, and more. Read more: nuanced.dev/blog/beyond-file…