Build from your specs and docs.The right context, fewer tokens.
Mimir indexes your specs, docs, and code locally and gives your AI agents only the useful cited passages, over MCP. Build from your requirements, without burning tokens on your whole repo.
Fewer tokens
Only the cited passages reach the model, not your whole repo.
100% local
Your specs, docs, and code stay on your machine.
Native MCP
Claude, Codex, Cline, and your agents query the same cited sources.
How it works
You run
Your specs, docs & code, indexed locally.
In Codex, you ask
Build the CSV export from spec-042.md
Mimir sends only the cited passages, not the whole repo.
Codex builds it, grounded in the sources
Grounded in your specs. Fewer tokens burned.
Use cases
What your agents and teams can do with Mimir.
Point Mimir at your specs, docs, and code. Your agents get the cited passages relevant to the task, not a token-heavy dump of the whole corpus.
Codebase & features
Understand and extend a codebase, fast.
Index code, runbooks, ADRs, READMEs, and decisions. The agent finds where a feature lives, what to read before coding, and ships changes with cited sources: onboarding a dev goes from weeks to minutes.
Audio and micro-learning
Listen to a dense dossier instead of reading it.
Generate a short audio summary (MP3/WAV) from cited passages with mimir audio, to review a spec, an architecture, or a watch without staying in front of the screen.
Refactor & migration
Change risky code with the full picture.
Before an agent refactors or migrates, Mimir gives it the cited callers, contracts, and past decisions: changes are planned on what the codebase actually does, not on a guess.
Debug & tracing
Trace a bug across the whole codebase.
Ask where a flow is implemented and why it fails; Mimir returns the cited files and passages along the call path, giving the agent the big picture it usually lacks.
Team memory
Answer without waking up the person who knows.
Team members query policies, decisions, incidents, and project notes from the same bounded local context.
Sales engineering
Answer RFPs and security questionnaires.
Retrieve evidence already written in security docs, contracts, SLAs, and compliance answers without copying it into a third-party SaaS.
Vendor due diligence
Explore a data room or a vendor dossier.
Analyze contracts, policies, technical appendices, and compliance documents with citations instead of scattered manual reading.
Legal and tax
Compare clauses, obligations, and evidence.
Contracts, tax notes, company memos, or residency research stay local while the agent retrieves the relevant sources.
Research and reports
Write from cited sources.
Briefings, Markdown reports, and summaries can separate evidence, inferences, uncertainties, and items to review.
Sovereign and private
The key strength: make documentation useful without exporting it.
Mimir keeps control at the project level: no telemetry, no automatic upload to a platform, no vague compliance promise, and explicit limits on what gets indexed.
Documents on disk
You point Mimir at the relevant folder. Private files stay in the repository or the local workspace.
Git-ignored state
The index, reports, and generated helpers live in the local git-ignored .mimir/ folder.
Redaction and explicit limits
Redaction, size limits, and auditing of unsupported files avoid claiming to index more than what is reliable.
Minimal logs
Access records stay metadata-oriented so usage can be checked without exposing sensitive content.
Install
One package to activate queryable local documentation.
Mimir starts as an open-source library and CLI. Install it where docs, runbooks, decisions, contracts, or questionnaires already live, then generate local state and agent helpers. The default mode stays offline, with no model to download.
Choose your package manager
Pick your package manager once. Every command below adapts to match.
Add Mimir to your project
Prepare the workspace
Connect useful agents
Search the evidence
Agent workflows
A local knowledge base for agents, onboarding, and teams.
Mimir generates per-agent MCP helpers from the same project state. You explicitly choose Claude, Codex, Kimi, OpenCode, Cline, or a stricter subset, without rebuilding a knowledge base per tool.
Agent command
Target clients
Each agent stays connected to the same local server bounded to the workspace.
Claude
Project skill and MCP helpers to retrieve local evidence during Claude Code sessions.
Codex
Codex helpers keep search bounded to the selected repository, not a global corpus.
Kimi
Kimi gets the same local MCP entry point to keep evidence consistent across agents.
OpenCode / Cline
OpenCode, Cline, and custom MCP clients can use the same generated local server contract.
Desktop client coming
Desktop will come after the open-source core, with direct download.
The Tauri client should make the local flow visible: pick projects, watch folders, and run the same Mimir commands without a terminal. No fake pricing page or waiting list until the native release is ready.
Follow the projectMimir Desktop
PlannedDistribution planned through direct downloads and sideloadable installers, not through the App Store or Play Store.
Project registry
Each local folder keeps its own Mimir state and its own access limits.
Direct distribution
Desktop installers and Android APK/sideload are the targeted paths before any store consideration.
Same local core
The app wraps Mimir instead of creating a second hosted index or cloud document storage.
FAQ
Frequently asked questions
What Mimir does, what it never sends off your machine, and how to connect it to your agents.
What is Mimir?
Mimir is an open-source library, CLI, and MCP server that turns any local codebase or document corpus into cited, queryable context for your AI agents. It indexes your files on your machine and returns cited passages: your agents answer from your real code and real documents instead of guessing. It works offline by default.
Does Mimir send my confidential code or documents?
No. Mimir only indexes the local folders you choose and keeps all generated state in the git-ignored .mimir/ folder. No telemetry, no upload to the cloud, and offline by default: it fits confidential code, customer data, and regulated environments.
Which AI agents and tools does Mimir work with?
Any MCP-compatible agent: Claude Code, Codex, Kimi, OpenCode, Cline, and custom MCP clients. Run mimir install-agent to generate the configuration and skills for each one, or use the CLI and the TypeScript library directly.
Does Mimir require an API key, a model download, or an internet connection?
No. The default local-hash mode runs entirely offline, with no model or API key. Optional semantic embeddings (Transformers.js) can be preloaded once for better search quality, then also run offline.
How does Mimir differ from a cloud RAG or an IDE's codebase indexing?
Hosted RAG and IDE indexing usually send your code chunks to their servers to compute embeddings. Mimir keeps everything on your machine, exposes it over MCP to any agent, and returns cited passages: local, cited, MCP, code and documents, with no vendor lock-in.
Does Mimir write the answers for me?
No. Mimir returns cited passages from your local files; your agent or model writes from those sources. By keeping generation out of the core, answers stay grounded in real evidence and every citation is verifiable.
Which file formats can Mimir index?
Source code, text, Markdown, PDF, Office and OpenDocument, EPUB, HTML, CSV, JSON, and YAML, plus the custom text extensions you enable. Unsupported files are flagged explicitly, and scanned PDFs can use an optional OCR command.
Is Mimir free and open-source?
Yes. Mimir is MIT-licensed and free, with no account required. The CLI, the library, and the MCP server are the current product; a desktop client is planned later as a direct download.
Start with the core
Install Mimir, keep the dossier local, query the evidence.
No hosted account is required, and no data is exported to a third-party service.