Skip to content
Graft Docs

Architecture Decisions

Local-first, zero infrastructure

Section titled “Local-first, zero infrastructure”

Graft makes deliberate architectural choices that distinguish it from other code intelligence tools:

No vector database

Section titled “No vector database”

Many context tools embed code into vectors and run similarity search. Graft uses linked Markdown files instead:

  • No index to keep warm. The graph is files, not a running process.
  • No similarity search. Agents grep and follow links — the same tools they use for any file.
  • No drift. A stale graph shows up as a diff in review, not as silently wrong search results.

No daemon

Section titled “No daemon”

Graft has no background process. The graph is rebuilt on demand:

  • graft build rebuilds explicitly
  • Query commands auto-refresh in ~3ms before answering
  • Claude Code hooks trigger rebuilds after each agent turn

Git-native

Section titled “Git-native”

The graph is just files in graft/:

  • Portable — Anyone who clones the repo has the graph (if committed) or regenerates it with graft build.
  • Diffable — When code changes move things around, the graph diff appears in the same PR.
  • No setup — No database, no server, no configuration beyond graft init.

Your provider, your key

Section titled “Your provider, your key”

Summaries are written by any provider you choose:

  • OpenAI, Anthropic, Google, Ollama, OpenRouter, LiteLLM proxy
  • Under your own API key
  • The structural graph (tree-sitter) never calls a model

No telemetry

Section titled “No telemetry”

Graft collects no usage data, sends no analytics, and phones home to nothing.

See also

Section titled “See also”