Wordkeep

An AI-agnostic Model Context Protocol server that turns large codebases into high-signal, budgeted context for coding agents and humans. Built for engine-scale repos where grep and full-file reads do not scale.

The problem

Agents burn context on noise: whole files, stale indexes, and hand-rolled prompts. Wordkeep answers with structured tools: symbol maps, call graphs, knowledge search, defect tracking, and Tracy triage. The model sees what matters.

What we shipped

• Rust MCP server with composable tools for repository mapping, symbol refs, call paths, test discovery, change-impact analysis, and session handoff.
• Multi-language analysis via tree-sitter (C/C++, Rust, Python, C#, TypeScript, Svelte, GLSL, and more) with incremental reparsing and mtime-based caches.
• Deterministic BM25 retrieval plus optional local semantic rerank; telemetry for latency, outcomes, and tokens saved.
• Dark plum wiki UI aligned with Arathyll's design tokens for notes, defects, and atlas content.

Results (measured on Betwixt)

In daily agent workflows on this engine repo, Wordkeep routinely cuts context payloads by 99%+ while keeping warm queries in the single-digit to low tens of milliseconds range; enough to stay on the critical path of interactive coding.

Public vs private

Production Wordkeep stays loopback-first: MCP over stdio in your editor, optional wiki GUI on 127.0.0.1:8787, no write API on the public internet.

This site hosts the Wordkeep Wiki lab: search, reader, MCP telemetry dashboard, and garden health over a read-only sanitized corpus. It is a portfolio demo, not a hosted Wordkeep instance.

See also

• Lab: Wordkeep Wiki

  Interactive browser demo over the sanitized corpus.

• GitHub

  Install, MCP config, and the wiki companion.

• Services

  How we embed similar discipline on client codebases.