These are the questions people ask first about OpenContext, answered directly. (Two terms recur: an API key is the credential that identifies your account with a model provider, and MCP is the Model Context Protocol, the standard chat clients use to call tools.) Every answer describes the code as it is today; the how it works, privacy, and for-agents pages have the detail behind each one.
What is OpenContext?
OpenContext is a local-first app that turns documents, email, repositories, and conversations into a knowledge graph of plain markdown files on your own machine, one entity per file with relationships as relative links. Coding agents and chat assistants read that graph as memory through the opencontext command-line tool and a Model Context Protocol (MCP) server, and contribute back through the same pipeline that screens and names everything.
Is OpenContext free?
Yes, during the closed beta. The beta is invite-only and ships under an evaluation license that covers the software only; your vault, your documents, and everything you create with it remain yours. Running it costs whatever your own model provider charges for the calls you trigger, and demo mode makes no model calls at all.
Does OpenContext train on my data?
No. OpenContext trains nothing and has no server to send your data to. Your content goes to exactly one place, the model provider you configured (Anthropic directly, or OpenRouter), under your own API key and your own agreement with that provider.
Where does my data live?
In a folder on your machine that you choose, called the vault, holding the knowledge graph as markdown files, its own git repository, the originals under _sources/, and a _state/ directory for settings and credentials. The local server listens only on the loopback address 127.0.0.1. Nothing is written anywhere else except a small file remembering which vault you chose and the optional agent-link block in repositories you linked.
Which models does it use?
The default is claude-sonnet-5 through Anthropic's Agent SDK (its software development kit), with an Anthropic API key, an OpenRouter key, or a logged-in Claude Code command-line tool. With OpenRouter selected you can pick any model OpenRouter serves; models that support tool use are recommended because vault-building leans on structured output. Demo mode runs a deterministic simulator instead of a model.
Does it work with Claude Code, Cursor, Codex, and Claude Desktop?
Yes. Agents that can run a shell, such as Claude Code and the Codex command-line tool, use the opencontext command-line tool, and opencontext link writes a primer into the repository's AGENTS.md, CLAUDE.md, CODEX.md, and README.md so the agent knows the graph exists. Claude Desktop, Cursor, Codex, and any other MCP client use the MCP server, with a ready-made install snippet per client in Settings, Connect your AI tools.
What is a context graph?
A context graph is the knowledge an agent needs about your world, stored as linked files rather than as a prompt. In OpenContext that is one markdown file per person, organization, project, meeting, decision, or concept, with frontmatter naming its type, name, aliases, and sources; relationships written as relative markdown links between files; a README index in every directory; and a source note recording where each fact came from. Because it is files, any agent, editor, or grep can read it, and git records every change.
How is this different from retrieval-augmented generation?
Retrieval-augmented generation (RAG) fetches passages of raw documents by embedding similarity at question time. OpenContext distills documents once into named entities with explicit relationships and provenance, and an agent retrieves by deterministic search over names and aliases, then follows the links between files. There is no embedding index, deliberately: ranked search over names and aliases is the design until it measurably underperforms. The graph is readable without any retrieval system at all, and an agent can traverse it from the root README like a wiki.
Can I export my context graph?
Yes, in two ways, both complete. Download my vault (Settings, Environment) streams a zip of every knowledge file, the _sources/ originals, and the full git history. Or push the vault to a GitHub or GitLab remote of your own, which carries the entities, every index, and a vault-root AGENTS.md primer. Secrets, logs, extracted text, and repository clones stay behind in both cases.
What is agent linking?
Agent linking is how a repository learns the graph exists. Installing the CLI, or running opencontext link, writes a managed block between <!-- opencontext:begin --> and <!-- opencontext:end --> markers into the repository's AGENTS.md, CLAUDE.md, CODEX.md, and README.md, naming the vault, the read commands, and that repository's own project graph. It is idempotent, works only inside a git worktree unless told otherwise, previews with --check, and opencontext unlink restores the files byte for byte.
Do I need an account?
No. There is no OpenContext account, no sign-up, and no OpenContext server. You need a model credential of your own (an Anthropic key, an OpenRouter key, or a Claude Code login), or none at all in demo mode. Connecting Google Drive, Gmail, GitHub, GitLab, or X uses your own sign-in grant (an OAuth client you register) or personal token, with read-only scopes.
Does it run on Windows?
Best effort. macOS and Linux are supported. On Windows, npm run dev, npm run demo, and npm start work because the scripts avoid shell-specific syntax, but npm run free-ports relies on lsof and does nothing there. The command-line tool's installer has a PowerShell one-liner.
What leaves my machine?
Only calls you asked for. Your model provider is the path that carries your content; the optional entity scan is the one agent that fetches public web pages from your machine; Google is called when you connect Drive or Gmail, GitHub or GitLab for accounts you connect and repositories you register, X for links you paste and bookmark syncs you press, a page you paste a link to is fetched once, and OpenRouter's model catalog is read while the Models tab is open. There is no telemetry, no analytics, and no account, and the web app loads no third-party scripts or fonts.
Can an agent write to the graph?
Only through the ingestion pipeline. opencontext contribute, opencontext capture, and the MCP contribute tool submit text that is deduplicated, screened for sensitive values before any model reads it, named by convention, and attributed in git history. No agent, including OpenContext's own, writes vault files directly; every write passes a path sandbox, a naming check, and a privacy screen.
What does it cost to run?
The software is free during the beta; the model calls are yours. Distillation is one toolless model call per document by default, estimated from early runs at roughly a tenth of the cost of the agent loop it replaced (a measured comparison is still owed), and search, the index audit, duplicate detection, the privacy audit, and export cost nothing because no model runs. Settings can set daily and weekly spend budgets, and every run can be stopped before or while it spends.