The archive · Developer & Business Tools · Product decision · 2026
Ataraxy Labs bets git diffs should read as entities; sem hits 3.3k stars
sem parses code with tree-sitter and diffs at entity level — functions, not lines — for humans and AI agents; 3.3k stars plus a published accuracy benchmark.
Ataraxy Labs
What the business is
sem is an open-source semantic version-control CLI that sits on top of git: it parses code with tree-sitter into entities (functions, methods, classes), then diff, blame, impact, log, entities and context commands report change at entity level, with JSON output and a token-budgeted context command aimed at LLM agents.
How it started
Ataraxy Labs created the repository on 2026-02-05 and positioned sem as 'semantic version control on top of git': tree-sitter parses code, extracts every function, class and method as an entity, and diffs at the entity level with rename detection, structural hashing and word-level inline highlights — 'function authenticateUser modified' instead of 'lines x-y changed.' Six commands cover diff, blame, impact, log, entities and token-budgeted context, all working in any git repo with no configuration.
What happened
The product site claims roughly 26 languages and 5 data formats via tree-sitter, 8ms typical diffs and 4,000+ downloads, and publishes an agent-accuracy benchmark: Claude Sonnet 4.5 answered the same code-change questions from sem JSON versus raw git diff across 3 commits, and sem's entity output was exact or accurate where git-based answers confused added lines with added entities (238 vs 32 on a speed-optimization commit) or counted files instead of entities. The repo README, live 2026-09-05, documents what came next: entity matching phases, a GitHub Action that posts entity-diff comments on pull requests, an MCP server with eight sem tools for coding agents, and an opt-in cloud-backed graph for large repositories and teams.
How it ended up
No company, funding or revenue is documented in the material; as of 2026-09-05 sem is live and shipping, with the repo grown from a six-command CLI into a GitHub Action, an MCP server and an opt-in cloud tier.
Background
sem is an open-source semantic version-control tool that runs on top of git. It parses code with tree-sitter, extracts functions, methods, classes and types as entities, and reports changes at that level — 'function authenticateUser modified' instead of 'lines x-y changed.' Its six commands cover entity diff, blame, impact analysis, entity history, entity listing and a token-budgeted context command built for LLM prompts, all working in any git repo with no configuration.
The bet is that git's line-level diff is the wrong lens for both code review and AI agents. The site claims roughly 26 languages and 5 data formats via tree-sitter, 8ms typical diffs and 4,000+ downloads, and publishes an agent benchmark: Claude Sonnet 4.5 answered the same questions from sem JSON and from raw git diff across 3 commits, and sem's answers matched ground truth where git-based answers counted added lines as added entities (238 vs 32) or files instead of entities.
Ataraxy Labs created the repo on 2026-02-05; at the 2026-09-04 crawl it showed 3.3k stars (3,334 in metadata). The live repo README adds the shape of what came next — a GitHub Action that posts entity diffs on pull requests, an MCP server with eight sem tools for coding agents, and an opt-in cloud graph for large monorepos and teams. No company, funding or revenue is documented, so the entry records a developer-tools bet whose measurable results so far are stars, downloads and a public benchmark.
What has to be true
- Concrete traction: Ataraxy-Labs/sem, created 2026-02-05, showed 3.3k stars (3,334 in crawl metadata) at the 2026-09-04 crawl, with 4,000+ downloads claimed on the product site.
- Falsifiable claim with published evidence: the site's '2.3x more accurate' line links to a benchmark showing sem JSON exact where git-diff answers were wrong (238 vs 32 added entities).
- Clear architectural bet: functions and entities as the unit of diff, blame, impact and log is a coherent answer to the line-diff failure modes the benchmark documents.
- Agent-native direction: JSON output, a token-budgeted context command and, per the repo, an MCP server all treat coding agents as first-class users.
What can be applied
Fix the unit of analysis first: line diffs mislead both reviewers and agents; once the parsed entity is the unit, accuracy improves downstream for humans and LLMs without changing the commands.
Aftermath
As of 2026-09-05 sem is live: the repo shows 3.3k stars, installs via Homebrew, winget, Scoop, npm, cargo or Docker, and the README documents entity-diff GitHub Actions, an MCP server with eight sem tools, and an opt-in 'sem cloud' graph for large repositories and teams. The material records no company entity, funding or revenue, so sem is best read as an Ataraxy Labs open-source developer-tools bet whose evidence so far is GitHub traction, downloads and a public agent-accuracy benchmark.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card