mimirs

MIMIRS

Named after Mímir, the Norse god of wisdom and knowledge.

Persistent project memory for AI coding agents. One command to set up, nothing to maintain.

Your agent starts every session blind — guessing filenames, grepping for keywords, burning context on irrelevant files, and forgetting everything you discussed yesterday.

On one real project, a typical prompt was burning 380K tokens and ~12 seconds end-to-end.

After indexing with mimirs: 91K tokens, ~3 seconds — a 76% drop on that codebase. Your numbers will vary with repo size, query, and model.

No API keys. No cloud. No Docker.
Just bun and SQLite.

Semantic Search  ·  Auto-generated Wiki
Cross-session Memory  ·  Dependency Graphs  ·  Annotations

Works with: Claude Code  ·  Cursor  ·  Windsurf  ·  JetBrains (Junie)  ·  GitHub Copilot  ·  any MCP client

Quick start

1. Prerequisites

Bun (curl -fsSL https://bun.sh/install | bash) and, on macOS, a modern SQLite — Apple's bundled one doesn't support extensions:

brew install sqlite

Linux and Windows ship with a compatible SQLite already.

2. Set up your editor (automatic)

bunx mimirs init --ide claude   # or: cursor, windsurf, copilot, jetbrains, all

This creates the MCP server config, editor rules, .mimirs/config.json, and .gitignore entry. Run with --ide all to set up every supported editor at once.

init covers Claude Code, Cursor, Windsurf, Copilot, and JetBrains (Junie). For everything else — Codex, Zed, custom clients — copy one of the snippets below.

3. Set up your editor (manual reference)

The mimirs MCP server runs over stdio. Every client needs the same three things: a command (bunx), args (["mimirs@latest", "serve"]), and a RAG_PROJECT_DIR env var pointing at your project root.

Claude Code — .mcp.json in project root

{
  "mcpServers": {
    "mimirs": {
      "command": "bunx",
      "args": ["mimirs@latest", "serve"],
      "env": {
        "RAG_PROJECT_DIR": "/absolute/path/to/your/project"
      }
    }
  }
}

Cursor — .cursor/mcp.json in project root

{
  "mcpServers": {
    "mimirs": {
      "command": "bunx",
      "args": ["mimirs@latest", "serve"],
      "env": {
        "RAG_PROJECT_DIR": "/absolute/path/to/your/project"
      }
    }
  }
}

Windsurf — ~/.codeium/windsurf/mcp_config.json (global)

Windsurf reads MCP servers from your home directory, not the project. JetBrains plugin variant uses ~/.codeium/mcp_config.json.

{
  "mcpServers": {
    "mimirs": {
      "command": "bunx",
      "args": ["mimirs@latest", "serve"],
      "env": {
        "RAG_PROJECT_DIR": "/absolute/path/to/your/project"
      }
    }
  }
}

JetBrains (Junie) — .junie/mcp.json in project root

{
  "mcpServers": {
    "mimirs": {
      "command": "bunx",
      "args": ["mimirs@latest", "serve"],
      "env": {
        "RAG_PROJECT_DIR": "/absolute/path/to/your/project"
      }
    }
  }
}

GitHub Copilot — .vscode/mcp.json in project root

VS Code's Copilot uses a servers map (not mcpServers) and a type field.

{
  "servers": {
    "mimirs": {
      "type": "stdio",
      "command": "bunx",
      "args": ["mimirs@latest", "serve"],
      "env": {
        "RAG_PROJECT_DIR": "/absolute/path/to/your/project"
      }
    }
  }
}

Codex — ~/.codex/config.toml (global)

Codex uses TOML, not JSON, and reads from ~/.codex/config.toml. One block per project — pick a unique table name if you wire up multiple repos (mimirs-frontend, mimirs-api, etc).

[mcp_servers.mimirs]
command = "bunx"
args = ["mimirs@latest", "serve"]
env = { RAG_PROJECT_DIR = "/absolute/path/to/your/project" }

Or, equivalently, with an expanded env table:

[mcp_servers.mimirs]
command = "bunx"
args = ["mimirs@latest", "serve"]

[mcp_servers.mimirs.env]
RAG_PROJECT_DIR = "/absolute/path/to/your/project"

Read-only project directory? Redirect the index

If the project lives in a read-only mount, set RAG_DB_DIR to a writable location. The index lives there instead of <project>/.mimirs/.

{
  "mcpServers": {
    "mimirs": {
      "command": "bunx",
      "args": ["mimirs@latest", "serve"],
      "env": {
        "RAG_PROJECT_DIR": "/read/only/project",
        "RAG_DB_DIR": "/home/me/.cache/mimirs/myproject"
      }
    }
  }
}

4. First index

The MCP server indexes lazily on the first query, so once it's wired up you can just ask your agent something. To force a full index up front (useful for large repos):

bunx mimirs index            # current directory
bunx mimirs status           # how many files, chunks, embeddings

5. Try the demo (optional)

bunx mimirs demo

Manual workflow (without init)

init is a convenience: it wires up your editor (MCP config, agent rules, .gitignore, .mimirs/config.json). It does not build the index, and nothing below needs it — the index and a default config are created automatically the first time you index or query.

1. Add the MCP server by hand. Drop the snippet for your client from the manual reference above: command: "bunx", args: ["mimirs@latest", "serve"], and RAG_PROJECT_DIR pointing at your project root. That is the entire MCP setup.

Without init there's no agent-rules file, so your assistant won't know the tools exist. Either mention mimirs in your prompt, or copy the tool list from CLAUDE.md into your editor's rules.

2. Build the index. The MCP server indexes lazily on the first tool call, so through an agent you can skip this step. To index up front (recommended for large repos, and required before the CLI search/read below):

bunx mimirs index                                # current directory
bunx mimirs index /path/to/repo                  # a specific directory
bunx mimirs index --patterns "src/**/*.ts,*.md"  # restrict to globs
bunx mimirs status                               # files, chunks, embeddings

No init and no config file required — defaults are applied and the index is written to <project>/.mimirs/.

3. Query from the CLI. Two read commands, both running against the index in the current directory (use --dir to point elsewhere):

# Where is it? — ranked file paths + snippet previews
bunx mimirs search "where is auth handled" --top 10

# What is it? — the actual matching code chunks (functions, classes, sections)
bunx mimirs read "jwt validation" --top 8 --threshold 0.3

Scope either with --ext .ts,.tsx, --in src,packages/core, or --exclude tests. Note: the CLI search/read do not auto-index — run mimirs index first (only the MCP server indexes on demand).

Claude Code plugin

For deeper integration, mimirs is also available as a Claude Code plugin. In a Claude Code session:

/plugin marketplace add https://github.com/TheWinci/mimirs.git
/plugin install mimirs

The plugin wires the MCP server, three hooks — SessionStart (context summary), PostToolUse (auto-reindex on edit), SessionEnd (auto-checkpoint) — and a set of workflow skills that orchestrate the tools for common jobs: explore, plan, review, debug, research, recall, catch-up, handoff, doc-gaps, scout, and wiki.

Want the skills without the plugin? They're plain SKILL.md files under skills/. Copy any you like into your project's .claude/skills/<name>/ (shared with the repo) or ~/.claude/skills/<name>/ (all your projects) and Claude Code picks them up next session. Skills are a Claude Code feature, so they don't apply to other editors — but the MCP tools themselves work everywhere.

Search quality

89–97% Recall@10, 97–100% Recall@20, MRR 0.69–0.77. Benchmarked on four real codebases across three languages with stratified, difficulty-mixed query sets (72–120 queries each, ~⅓ hard), re-measured 2026-06-04 on the current pipeline. Full methodology in BENCHMARKS.md.

Codebase	Language	Files	Queries	Recall@10	MRR	Zero-miss
mimirs	TypeScript	244	74	95.3%	0.759	4.1%
Excalidraw	TypeScript	693	72	90.3%	0.773	9.7%
Django	Python	3,181	116	97.4%	0.727	2.6%
Kubernetes	Go	8,792	120	89.2%	0.689	10.8%

The larger repos (Kubernetes, Excalidraw) are big enough that some correct files rank just past the top-10; recall reaches 97–100% by top-20, so set searchTopK: 15–20 on large repos.

vs coding agents (ContextBench)

We also ran mimirs on ContextBench (gold-context retrieval on real repos), whose other entries are full coding agents — multi-step explorers — not single-call tools. Given a focused query (what an LLM sends after reading the issue), one mimirs retrieval call ranks like this against whole agent trajectories:

metric	mimirs	rank	field
File coverage	0.799	#1	above OpenHands, SWE-agent, Agentless…
Line coverage	0.341	#1	above Agentless, mini-SWE…
Line precision	0.316	#2	behind only Agentless (0.376)
File precision	0.192	#6	low by design — recall-first

mimirs leads both coverage metrics as a single call. File precision is last on purpose: a missed gold file is fatal (the LLM never sees the code to fix), an extra file reference is cheap to filter — so mimirs maximizes recall and lets the model do the precision pass. And that low file precision is mostly an artifact of the metric: ~86% of the non-gold files mimirs returns are relevant context coupled to the fix (callers, types, sibling implementations), not noise — measured against gold the precision is 0.19, against relevance it is 0.87.

Same recall, a fraction of the cost. Head-to-head against a grep-only agent (raw issue, no index, no peeking at the fix) localizing the same 15 issues: mimirs delivers the relevant cluster in one ~15 ms call with zero LLM tokens; the agent took ~11.5 tool calls per issue (each an LLM step) to converge — and stopped at the primary file. On multi-file fixes the agent reached 22% of the gold files, mimirs 56% in that single call — the dependency graph surfaces the secondary files the issue never names.

n=15 sample vs the agents' 500-set — directional; agent tool-calls self-reported and capped. Full leaderboards, caveats, relevance + cost tables in BENCHMARKS.md.

How it compares

	mimirs	No tool (grep + Read)	Context stuffing	Cloud RAG services
Setup	One command	Nothing	Nothing	API keys, accounts
Token cost	~91K/prompt	~380K/prompt	Entire codebase	Varies
Search quality	89–97% Recall@10	Depends on keywords	N/A (everything loaded)	Varies
Code understanding	AST-aware (24 langs)	Line-level	None	Usually line-level
Cross-session memory	Conversations + checkpoints	None	None	Some
Privacy	Fully local	Local	Local	Data leaves your machine
Price	Free	Free	High token bills	$10-50/mo + tokens

Why not an existing tool?

• Continue.dev's @codebase — closest overlap (local RAG, open source), but retrieval lives inside the editor extension. Mimirs is a standalone MCP server with explicit tools (search, read_relevant, project_map, search_conversation, annotate) the agent can plan around, plus conversation tailing and a wiki generator built in.
• Aider's repo-map — static tree-sitter summary of the repo, no embeddings. Clever and lightweight, but a summary isn't retrieval — mimirs ranks chunks per query with vector + BM25 and boosts by graph centrality.
• Sourcegraph Cody / OpenCtx — excellent at code search, but indexing leans on cloud infra and an account. Mimirs is one bunx away and never leaves your machine.
• llama-index / LangChain / roll-your-own — those are libraries. Mimirs is batteries-included: AST-aware chunking, hybrid retrieval, file watcher, conversation tail, and annotations already wired together.

How it works

• Parse & chunk — Splits content using type-matched strategies: function/class boundaries for code (via tree-sitter across 24 languages), headings for markdown, top-level keys for YAML/JSON. Chunks that exceed the embedding model's token limit are windowed and merged.

• Embed — Each chunk becomes a 384-dimensional vector using all-MiniLM-L6-v2 (in-process via Transformers.js + ONNX, no API calls). Vectors are stored in sqlite-vec.

• Build dependency graph — Import specifiers and exported symbols are captured during AST chunking, then resolved to build a file-level dependency graph and a symbol-level call graph. impact walks the transitive callers of a function (blast radius + tests to run); trace finds how one symbol reaches another; the mimirs affected CLI turns a git diff into the exact set of tests to run.

• Hybrid search — Queries run vector similarity and BM25 in parallel, combined by reciprocal-rank fusion (weighted, default 0.5) — robust to the two scorers' very different score scales. Identifiers are split (camelCase/snake_case) so a search for depends matches getDependsOn. Results are then boosted by dependency graph centrality and path heuristics. read_relevant returns individual chunks with entity names and exact line ranges (path:start-end).

• Watch & re-index — File changes are detected with a 2-second debounce. Changed files are re-indexed; deleted files are pruned.

• Conversation & checkpoints — Tails Claude Code's JSONL transcripts in real time. Agents can create checkpoints at important moments for future sessions to search.

• Annotations — Notes attached to files or symbols surface as [NOTE] blocks inline in read_relevant results.

• Analytics — Every query is logged. Analytics surface zero-result queries, low-relevance queries, and period-over-period trends.

Data handling

mimirs runs entirely on your machine. It indexes files your repo tracks plus untracked-but-not-gitignored files (so a .env you forgot to gitignore could be read — common secret patterns like .env, *.pem, *.key, and SSH keys are excluded by default; add your own to exclude in .mimirs/config.json). File content and embeddings are stored in <project>/.mimirs/index.db, a local SQLite file. Conversation indexing reads only the current project's transcripts under ~/.claude/projects/<this-project>/.

The only network call is a one-time download of the embedding model (Xenova/all-MiniLM-L6-v2) from huggingface.co, cached at ~/.cache/mimirs/models. Your code never leaves your machine — nothing is sent to any server.

Supported languages

AST-aware chunking via bun-chunk with tree-sitter grammars:

TypeScript, JavaScript, Python, Go, Rust, Java, C, C++, C#, Ruby, PHP, Scala, Kotlin, Lua, Zig, Elixir, Haskell, OCaml, Dart, Bash/Zsh, TOML, YAML, HTML, CSS/SCSS/LESS

Also indexes: Markdown, JSON, XML, SQL, GraphQL, Protobuf, Terraform, Dockerfiles, Makefiles, and more. Files without a known extension fall back to paragraph splitting.

Documentation

• Example tool outputs — what your agent actually receives over MCP
• MCP tools, CLI & analytics
• Configuration & examples
• Benchmarks

Stack

Layer	Choice
Runtime	Bun (built-in SQLite, fast TS)
AST chunking	bun-chunk — tree-sitter grammars for 24 languages
Embeddings	Transformers.js + ONNX (in-process, no daemon)
Embedding model	all-MiniLM-L6-v2 (~23MB, 384 dimensions) — configurable
Vector store	sqlite-vec (single .db file)
MCP	@modelcontextprotocol/sdk (stdio transport)
Plugin	Claude Code plugin with skills + hooks

All data lives in .mimirs/ inside your project — add it to .gitignore.