Graft vs ontology-atlas
Context layer for large codebases: a graph of plain-English markdown nodes — no embeddings, no index — agents read like any repo file. Claude Code hooks + MCP; 42% fewer tokens in its bench. — versus — Desktop app plus MCP server keeping a codebase's ontology as reviewable Markdown in the repo: domains and capabilities linked to code evidence, compiled to a graph for blast-radius queries.
Both hand agents a hand-written Markdown context layer over a large codebase instead of embeddings. Graft is a flat plain-English node graph with Claude Code hooks; Atlas constrains you to five node kinds and a fixed relation set so the folder becomes computable.
| Graft | ontology-atlas | |
|---|---|---|
| Stars | 5.2k | 85 |
| Forks | 467 | 14 |
| Language | TypeScript | TypeScript |
| License | MIT | MIT |
| Last activity | today | today |
| Topics | code-intel, rag | knowledge-graphs, code-intel, coding |
| Curated connections | 5 | 5 |
Graft — the curator's take
The anti-embeddings bet is the story: the codebase map is prose files — senior-engineer explanations, linked and greppable — so agents consume it with the file tools they already have, and a stale map shows up as a diff in review instead of rotting in a vector store. The 162-run controlled benchmark holding correctness equal is more rigor than this category usually shows; still self-run. Claude Code gets the deep integration (hooks, statusline, background rebuild), everyone else gets MCP. When NOT: nodes are LLM-written prose — building them on a big repo costs real tokens, and prose can drift subtly wrong in ways a symbol graph can't; spot-check the nodes on your hot paths. ★437 and young, with NanoNets behind it.
ontology-atlas — the curator's take
Pick it when the expensive question is *why* the code is shaped this way and what a change will touch. Because the ontology is authored and reviewed rather than inferred, unknown coverage shows up as unknown instead of a confident wrong graph — and every agent write lands as a diffable line in a Markdown file. It is not a code indexer: it will not find symbols or callers, and it only knows what someone wrote down, so an unmaintained atlas/ rots quietly while still looking authoritative. Practical caveat: this is a desktop install with an unsigned Windows beta; Linux and CI users run the CLI and MCP server from source.