Mimir – open-source code intelligence for AI agents (Go, MCP, SQLite)

Name: Mimir – open-source code intelligence for AI agents (Go, MCP, SQLite)
Author: thuongne

by thuongne·Mar 18, 2026·2 points·0 comments

View on HN

AI Analysis

●●●BangerBig BrainWizardry

Knowledge graph beats vector search for understanding call chains and blast radius.

Strengths

•Typed knowledge graph with CALLS/IMPORTS/EXTENDS edges enables structural reasoning.
•Go rewrite delivers 5x faster indexing and 17x smaller binary than Node predecessor.
•MCP server auto-configures into Cursor, Claude Code, VS Code, and Windsurf.

Weaknesses

•Only 8 languages supported limits adoption for polyglot codebases.
•Requires teams to adopt MCP ecosystem which is still early-stage.

Post Description

I built Mimir because every AI coding assistant I used was effectively doing regex search over my codebase. They had no understanding of call chains, module boundaries, or blast radius.

Mimir indexes a repo into a typed knowledge graph (nodes = symbols, edges = CALLS/IMPORTS/EXTENDS with confidence scores) and exposes it via MCP so agents can actually reason about code structure.

How it works: 1. mimir analyze . — parallel AST parse (go-tree-sitter, 8 languages), scope-aware cross-file resolution, Louvain community detection, BM25 + HNSW vector search 2. mimir daemon start — background MCP server 3. mimir setup — auto-writes config into Claude Code, VS Code, Cursor, Windsurf, Zed

Stack: Go 1.22, modernc/sqlite (pure Go, no CGO), sqlite-vec (HNSW), Sigma.js web UI

Benchmarks vs the Node.js predecessor: - Cold index 1k files: 45s → < 8s - Incremental re-index: 45s → < 2s - Binary size: 350MB → < 20MB - RAM during index: 800MB → < 120MB

GitHub: github.com/thuongh2/git-mimir Install: go install github.com/thuongh2/git-mimir/cmd/mimir@latest

Happy to answer questions about the architecture (resolver phases, incremental patching, the MCP transport layer).