mira

Mira

A local MCP server that gives Claude Code persistent context and code intelligence.

Mira runs alongside Claude Code over stdio. It tracks sessions, indexes your codebase with tree-sitter, and injects relevant context on every prompt via lifecycle hooks. Decisions, preferences, and code structure persist in SQLite across sessions. No cloud services, no accounts -- one binary, two database files.

Install

Claude Code:

claude plugin install mira
mira setup  # optional: configure API keys for semantic search

Gemini CLI:

gemini mcp add mira --command mira-wrapper --args serve
gemini skills link /path/to/mira/plugin/skills

Start a new session. Mira begins injecting context automatically via lifecycle hooks.

See docs/INSTALLATION.md for alternative methods (cargo install, manual binary, build from source).

What Mira Does

• Session continuity. Decisions, preferences, and context persist across sessions and surface automatically on every prompt. Stored facts are evidence-based -- they earn trust through repeated cross-session use, not blind storage.
• Semantic code search. "Where do we handle auth?" finds verify_credentials in middleware.rs, even when the word "auth" never appears. Falls back to keyword search without API keys.
• Background analysis. Indexes your codebase with tree-sitter, detects unused functions, doc gaps, and error patterns without being asked. Learns how errors were fixed and surfaces solutions in future sessions.
• Change intelligence. Diff analysis with call graph impact tracing and cross-session change tracking.
• Agent team coordination. File ownership tracking and conflict detection across concurrent agents. Agent team definitions for expert review, QA hardening, and safe refactoring.
• Goal tracking. Multi-session objectives with weighted milestones and automatic progress updates.

Design Principles

• Local-first. Two SQLite databases in ~/.mira/. No cloud services, no accounts, no external databases required.
• Evidence-based. Stored facts start as candidates and are promoted through cross-session use. What surfaces in your session is traceable, not just whatever was written last.
• Zero-config defaults. Context persistence, code intelligence, goal tracking, and background analysis all work without API keys. Add OpenAI embeddings for semantic search when you want it.
• Honest tooling. Context injection is conservative -- tight relevance thresholds, cross-prompt deduplication, suppression of signals that aren't being used. Mira tells Claude what it actually knows.

How It Works

Claude Code  <--MCP (stdio)-->  Mira  <-->  SQLite + sqlite-vec
    |                             |
    +--lifecycle hooks            +--->  OpenAI (embeddings, optional)

Mira runs as a local process spawned by Claude Code over stdio. Two databases: one for sessions, goals, and memories; one for the code index. Lifecycle hooks capture context at key moments (session start, prompt submit, tool use, compaction, stop) and inject relevant information back automatically.

Quick Start

After installing, Mira's slash commands are available natively in Claude Code and Gemini CLI:

/mira:recap           -- session context, preferences, and active goals
/mira:search <query>  -- semantic code search
/mira:goals           -- manage cross-session goals and milestones
/mira:diff            -- semantic analysis of recent changes
/mira:insights        -- surface background analysis
/mira:help            -- full command list

Capabilities With and Without API Keys

Capability	Without API Keys	With OpenAI Key
Memory and recall	Keyword/fuzzy search	Semantic search
Code search	FTS5 + fuzzy matching	Hybrid semantic + keyword
Code intelligence	Tree-sitter symbols, call graph	Same
Background analysis	Heuristic pattern detection	Same
Goal tracking	Full	Full
Agent coordination	Full	Full

OpenAI embeddings use text-embedding-3-small (~$1/month for typical usage).

Documentation