Your AI agent re-reads your whole codebase on every task. Standardoc indexes it once into a living map of your code — so the agent just asks. ~100 tokens per question instead of 30k. Local, open-source.
📖 English · Français | Quickstart · Roadmap · Changelog
Every task, your agent starts from zero: it greps, it reads files, it burns 30k tokens rebuilding context it already had last session. The bigger the codebase, the worse it gets — more tokens, more drift, more code that looks like yours but quietly breaks your invariants.
It reads your code straight from the syntax tree and keeps a living graph of it — every symbol, and the typed links between them: who calls who, what imports what, what implements what. A file watcher keeps it current as you type.
Your tools query that one graph instead of each re-parsing your code:
- Agents ask over MCP (
find_symbol,get_context,find_call_sites, …) — ~100 tokens where grep + read cost 30k. Claude Code, Cursor, Continue, Copilot, any MCP client. - Editors connect over LSP — the VSCode extension is built in; Neovim, Helix, JetBrains point at the same binary.
Rust, TypeScript / JavaScript (React, JSX, TSX), Vue, Svelte, Lua, and C today.
VSCode — search Standardoc in the Marketplace or Open VSX.
CLI (any agent, no VSCode):
cargo install --git https://github.com/miralabs-tech/standardoc standardoc-cli
standardoc init # wires the agent skill, MCP-first hooks, AGENTS.md, .mcp.jsonBig, complex codebases — compilers, engines, heavy monorepos. On a weekend
project ripgrep + your IDE are plenty; Standardoc earns its keep once the
archaeology starts costing you real time.
- One graph, not N. Every tool re-parsing your code is one more thing that drifts out of sync.
- Real AST, never regex. Heuristics rot the moment the code moves.
- Yours, for good. Local, FSL-1.1-MIT auto-converting to plain MIT (first release: April 26, 2028). No cloud, no lock-in, no rented graph.
Built by StandarX · Star the repo · Sponsor · Security policy
