max-context

Maximum Context

Maximum Context is a Go-based MCP (Model Context Protocol) server that gives AI coding assistants deep, always-current awareness of a codebase. It indexes a project's structure, symbols, dependencies, and architecture, then exposes that knowledge through a focused, decisive MCP tool interface.

Why max-context

AI coding agents waste 30–60% of their context window grepping and re-reading files to figure out where things are. max-context pre-computes that understanding and serves it on demand.

• Always current — file changes reflected in < 2 seconds via OS-native watchers
• No LLM at index time — deterministic tree-sitter parsing; no API keys, no token spend on indexing
• Minimal surface, no breaking changes — a small set of focused MCP tools, stable schema

See BENCHMARK.md for per-tool-call token savings vs naive and skilled Grep+Read baselines. The current skilled-baseline headline is 29.5× fewer tokens per deterministic query on max-context's own codebase.

For agent-session quality and cost — does an LLM actually do better with these tools? — see the causal A/B in experiments/eval/FINDINGS.md (same model, same tasks, only the toolset differs; blind different-model judge; pre-registered). On the tasks measured so far it finds equal answer quality to a skilled grep agent, at materially lower cost — up to 94% fewer tokens (≈90% in aggregate) on call-graph / "what-calls-this" questions — plus a structural edge on aliased imports (from m import f as g; g()), where text search can't connect the call site to the definition and max-context can. Honest scope: small sample (1 repo family, 1 replicate); quality is a tie, not a win, except the alias case.

How it works

                                  ┌─────────────────────────────────┐
                                  │       max-context binary        │
    ┌─────────────┐ fsnotify      │   ┌─────────────────────────┐   │
    │ Codebase    │───────────────┼─▶ │   tree-sitter parser    │   │
    └─────────────┘               │   └────────────┬────────────┘   │
                                  │                ▼                │
                                  │   ┌─────────────────────────┐   │
                                  │   │SQLite FTS5 + call graph │   │
                                  │   └────────────┬────────────┘   │
    ┌─────────────┐               │                ▼                │
    │ AI CLI / IDE│ stdio MCP     │   ┌─────────────────────────┐   │
    └─────────────┘◀──────────────┼───│  MCP server · 5 tools   │   │
                                  │   └─────────────────────────┘   │
                                  └─────────────────────────────────┘

Tree-sitter parses every supported language deterministically. Symbols and call edges land in a SQLite database with FTS5 indexes on function/type names. Recursive CTEs power call-chain and impact queries. No LLM is involved at index time — the host CLI's model does all the reasoning, with max-context feeding it pre-computed structure.

Features

• Focused MCP tools: get_definition (exact-name, canonical definition), query_codebase (BM25 fuzzy symbol search with explicit next action), get_call_chain (recursive caller/callee traversal), get_impact (change blast radius), get_architecture (pre-computed project summary)
• Real-time index: File watcher keeps the index current within 2 seconds of changes
• Multi-language: TypeScript, JavaScript, Python, Go, Rust, Java (and more via Tree-sitter). Full call-graph resolution for Go, TypeScript/TSX, and Python; other languages are parsed with name-based call resolution — see docs/LANGUAGES.md for the honest parsed-vs-resolved matrix.
• Universal CLI support: One max-context setup <cli> configures Claude Code, VS Code Copilot, Codex CLI, Antigravity, Cursor, and Windsurf

Install

Pick one: