OpenKnowledge is Inkeep’s bid to fold a local-first markdown vault into the loop that Claude Code, Cursor, Codex, and now OpenCode already run in. The product page pitches a “beautiful, AI-native markdown editor” that is private, open source, free, and local-first. The editor is not the interesting part. The interesting part is that the vault living as plain files on your disk is also exposed as a Model Context Protocol server that coding agents read from and write to, which is where “local-first” stops meaning “model-free.”
What OpenKnowledge actually is
OpenKnowledge is a WYSIWYG markdown editor that Inkeep positions against Notion and Google Docs, with markdown kept as the underlying storage format rather than a database blob. Per the product page, it ships as a macOS desktop app or as a CLI-served web app on Linux, Windows, and Intel Macs, installed via npm install -g @inkeep/open-knowledge. The repo README describes it as a “local-first markdown editor and LLM wiki with integrations for Claude, Codex, and other harnesses,” licensed under GPL-3.0-or-later.
The notable design decision is the MCP server baked into the product. Where a traditional wiki asks you to export, copy-paste, or sync to get content into a coding agent, OpenKnowledge exposes vault files as MCP tools so an agent can read context and write edits back in place. The product page frames three use cases: an “agent brain” that functions as persistent memory across sessions, engineering specs and runbooks co-edited by humans and coding agents, and a team knowledge base synced through git.
Where “local-first” and “AI-native” stop agreeing
The tension is straightforward once you trace it. “Local-first” is a storage claim: your notes live as plain markdown files on a machine you control, not in a vendor database. That part holds; the docs are explicit that files are “plain files you own,” platform-agnostic, and that the app “runs on your machine; private by design.”
“AI-native” is an access claim, and it points the other direction. The built-in MCP server hands the vault to Claude, Cursor, Codex, and OpenCode. Those agents are, in practice, cloud LLMs. When Cursor pulls a runbook into context to draft a fix, or Codex reads a spec to implement it, the vault’s contents travel to whichever model backs the agent. Local-first storage does not equal local-only inference. The threat model OpenKnowledge inherits is therefore not Obsidian’s.
The distinction matters because the marketing surface reads one way to a buyer skimming the hero, and the architecture reads another way to anyone who follows the MCP path to its endpoint.
What leaves the disk, and what stays
Two things to separate: the embedding index, and the agent’s context window.
For embeddings, OpenKnowledge does not document where its vector index lives in the reviewed pages; the docs emphasize the editor and the MCP surface rather than an embedding-store spec. The closest published analog is the Obsidian ecosystem, where Smart Connections keeps its index on-disk inside .smart-env/ in the vault, and pairing it with a local Ollama backend (nomic-embed-text or mxbai-embed-large) keeps the chat path on the box too. That is a configuration where the entire retrieval stack, embeddings and inference, never touches a vendor API.
For the context window, the path is clearer. An MCP read from Claude Code, Cursor, Codex, or OpenCode serializes vault content into the agent’s prompt, served by that agent’s backing model. That is the off-disk surface, and it is the design rather than a leak: the product page lists “no copy-paste, no export” as a feature.
The honest framing is that storage is local, the file format is portable markdown, but the moment an agent touches the vault the content is in a cloud model’s context. Whether that is acceptable depends on what is in your vault.
How it compares to Obsidian-plus-Ollama and Notion
The practitioner question is what each option gives you on the privacy axis, because the comparison charts tend to rank on price and model selection rather than on where bytes land.
A September 2025 four-way comparison of Notion AI, Obsidian AI, Mem, and Reflect ranks Obsidian highest on privacy (local files), puts Mem and Reflect as end-to-end encrypted, and Notion as cloud storage. Only Obsidian offers full offline AI in that comparison, and notably Reflect exposes an MCP server for Claude Code, Codex, and Cursor, which is the same integration shape OpenKnowledge now pitches.
Obsidian’s local-AI stack, Smart Connections plus a local Ollama model, is the configuration that keeps the whole retrieval-and-inference path on-box, with the index in .smart-env/ and chat served by a local model (per the Obsidian local-LLM guide). Notion is the opposite anchor: cloud storage, cloud models, no pretense of local-first.
OpenKnowledge sits in a third position. It keeps Obsidian’s markdown-folder exit, files on disk in a portable format, but adopts Notion’s reliance on cloud models for the AI features by routing through MCP to whatever agent you point at it. The trade is real agent ergonomics, a single editing surface humans and agents share, in exchange for vault content reaching a cloud LLM whenever an agent reads or writes.
The markdown-folder exit question is whether you can walk away. With OpenKnowledge, the underlying files are plain markdown, so a future migration is a directory copy. Real lock-in is the workflow, not the format: once specs, runbooks, and agent memory live in a vault that coding agents depend on, moving the vault means rewiring every agent’s MCP config. That friction is the actual switching cost.
License, maturity, and the name-collision caveats
Three things to check before echoing the “open source” headline.
First, the license. The OpenKnowledge repo is GPL-3.0-or-later, which is OSI-approved, so the open-source claim holds for this product. The caveat is that Inkeep’s sister project, the inkeep/agents framework, is licensed under Elastic License 2.0, a source-available “fair-code” license that is explicitly not OSI-approved and restricts certain competitive uses. Anyone evaluating the Inkeep ecosystem should not assume the license on one repo extends to the other.
Second, maturity. OpenKnowledge shipped six beta releases between June 25 and 26, 2026, v0.19.0 through v0.19.1-beta.6, per the release history, including a brand rename from “Open Knowledge” to “OpenKnowledge” (one word) in v0.19.1-beta.2 and the addition of OpenCode as a supported agent in v0.19.1-beta.5. That is daily-cadence beta software. The Q3 roadmap on the product page names collaborative editing of docs and spaces, one-click sharing, and git-backed sync, none of which is shipped yet; collaborative editing in particular is explicitly future work. Peg any feature claim to a specific beta version, because the cadence will make any snapshot stale within weeks.
Third, name collision. “OpenKnowledge” collides with at least two other things: Open Knowledge Maps, an unrelated non-profit that describes itself as the world’s largest AI-based scientific-knowledge search engine, and Google’s “Open Knowledge Format.” None of these are related to Inkeep’s product. The markdown editor and the scientific-search non-profit are easy to conflate in search and in citation.
Who should care, and who should wait
For a team already standardizing on Claude Code, Cursor, or Codex and writing specs in markdown, OpenKnowledge removes the friction of keeping agent context and human docs in sync, and the GPL-3.0 license plus plain-markdown storage means the file-format exit is intact. The cost is that vault content reaches whichever cloud LLM backs those agents, paid every time an agent reads the vault.
For anyone whose reason for using Obsidian is keeping notes off any model, the Obsidian-plus-Ollama stack is the configuration that preserves that property. OpenKnowledge optimizes for agent collaboration and accepts cloud-model exposure as the price. The two are solving different problems, and the “local-first” label obscures that the problems differ.
Inkeep itself is worth context for sizing the bet. The company (YC W23, founders Nick Gomez and Robert Tran) raised a $13M seed led by Great Point Ventures, Khosla Ventures, and Y Combinator, with Anthropic and Midjourney named as customers, after entering the market as a “Copilot for Support” powering Ask AI buttons in vendor docs, per its earlier Launch HN. Shipping a knowledge-base product is a logical move for a company whose original product was a docs-grounded agent. It also means OpenKnowledge is a bet from a company still in beta on the category, backed by investors whose thesis is agent infrastructure, not note-taking.
Frequently Asked Questions
What did the “Open Knowledge” to “OpenKnowledge” rename actually change?
The v0.19.1-beta.2 rename was cosmetic. The npm package stayed @inkeep/open-knowledge, the macOS appId stayed com.inkeep.open-knowledge, the deep-link scheme stayed openknowledge://, and the domain stayed openknowledge.ai, so existing installs and links kept working.
How does OpenKnowledge register itself with each coding agent?
Each agent gets the MCP server entry written into its own config file rather than through a global hook. In v0.19.1-beta.5, the OpenCode integration writes to opencode.json and the agent surfaces as ‘OpenCode (CLI)’ in a docked-terminal launcher alongside Claude Code, Codex, and Cursor.
How does OpenKnowledge relate to Inkeep’s earlier support-agent product?
Inkeep’s first product embedded an Ask AI widget in third-party documentation, built on the Vercel AI SDK with Claude Sonnet 3.5. OpenKnowledge inverts that surface: instead of grounding an agent in someone else’s docs, it exposes the user’s own markdown vault as the context the agent calls into.
What does git-backed sync on the roadmap add to the off-disk exposure?
Git-backed sync creates a second off-disk path independent of MCP. Once vault content is pushed to a git remote, it leaves the local disk whether or not any coding agent ever reads it, which ties the privacy model to wherever that remote is hosted.
How do you tell Inkeep’s OpenKnowledge apart from the scientific-search non-profit?
Open Knowledge Maps at openknowledgemaps.org is an unrelated non-profit scientific-knowledge search engine, and Google also uses an “Open Knowledge Format.” For Inkeep’s markdown editor, the disambiguating identifiers are the openknowledge:// deep-link scheme and the openknowledge.ai domain.
