Local-first knowledge graph
Shared context for your team and your agents.
Ingest a document once. Everyone on the team, and every AI agent they use, can traverse it, with the sources attached.
Desktop app · coming soon · your knowledge stays on your computer
Local checks for sensitive data. Your choice of AI provider. Privacy, explained
Swipe to scroll the page. Use two fingers to zoom or move the graph. Tap an entity to explore.
How it works
Your knowledge, connected in three steps.
Drop in what you already have
PDFs, decks, spreadsheets, email, Drive, Gmail, GitHub repos, exported chats. Screened for sensitive data before any model reads it.
An agent builds the graph, once
People, projects, decisions, meetings, and concepts each become one markdown file. Relationships are plain links. Every fact keeps its source.
linked → projects/relay.md
updated people/README.md
Answers for people. Context for agents.
Teammates chat with the graph and get cited answers. Coding agents and assistants read the same folder over CLI or MCP. Nobody ingests the file twice.
Early results
OpenContext reduces token costs, API calls, and tool calls.
20.3% fewer model API responses. 16.4% fewer tool calls.
Preliminary results. Results may vary.
Based on preliminary testing on one coding task, codebase, and model. Cost and token figures average three paired comparisons; API response and tool-call figures average the two pairs with recorded counts. Costs combine recorded and transcript-estimated usage and exclude graph creation and maintenance. Quality was mixed across these preliminary comparisons.
Read the full analysis →Get OpenContext on your computer.
OpenContext runs on your own computer. Desktop installers are coming soon.
Download NowTry it on your own questions
Stop asking around. Ask the graph.
Install openContext on your computer, add a source, and ask a question grounded in your own files.
Bring your own agent
Install with your agent.
Paste one prompt into Claude Code, Codex, or your local agent. Set up your context vault and preview a document scan, all from your conversation.
Local beta access requires an authorized checkout or an installation package.
Help me install OpenContext on this computer and create <username>-Context-Vault using my preferred username.
Read https://getopencontext.com/install-agent.md, https://getopencontext.com/agent-install.json, and https://getopencontext.com/for-agents.md first. Check current availability. Use the verified full local runtime, not a client-only MCP package or the unrelated public npm package named opencontext. If no public download is available, use an authorized checkout or ask me for access or the maintainer-supplied package. Do not invent a download URL.
Follow the installation guide, check prerequisites and any published checksum, preserve existing vaults, and connect your Model Context Protocol (MCP) client or command-line tools. Preserve existing client configuration and verify the connection. Complete setup without requiring the frontend.
Preview a metadata-only scan of my home directory for documents and useful hidden context. Exclude applications, system folders, credentials, caches, and dependencies. Include application folders or other locations only when I explicitly request them.
Show me the scope, exclusions, incomplete coverage, document count, token estimate, time range, and cost estimate where pricing is known. Explain the actual billing provider and any unknown subscription allowance. If subscription execution is unsupported, say so and do not switch to paid provider execution.
Wait for my approval of the scope and estimate before ingestion or paid model calls. Then use the existing privacy reviews and deterministic graph-writing pipeline. Report progress and bring privacy holds back to me for review. Keep my vault local; offer the frontend as an optional view of the same vault.