OpenKnowledge review MCP writes, search and restart recovery

Running published 0.89.3 against artificial Markdown to verify real MCP writes, edits, lexical search, ignore rules, file watching and persistence across server processes.

AgentIndex · Published 2026-11-07 · Updated 2026-11-07

OpenKnowledge connects a local Markdown workspace with an editor and agent tools. A foundational adoption question is whether an agent's successful save becomes a real file that can be searched and recovered after restarting the service. We tested that chain using the published HTTP/MCP server.

Version 0.89.3 completed writing, targeted editing, disk readback and restart recovery. A newly created external file became searchable. An edit whose target text was absent was explicitly refused without changing the document. These findings support a local document integration use case; they do not establish AI writing quality, visual editing fidelity or team-sync reliability.

A published package, with matching release source

On November 7, 2026, we installed @inkeep/[email protected] and executed its distributed dist/cli.mjs. Source inspection used release tag v0.89.3, commit 2a5c081bda0d0dedcccee4ae1f7ddd959032fbc0. Records include npm integrity and the installed entry-file hash. We did not attribute newer main-branch behavior to this release. Pinned release source

The host was macOS 26.6.2 with Node.js 26.3.0 ARM64, meeting the package's Node 24+ requirement. Initialization used --no-mcp --no-skills --content-dir notes; it skipped personal editor registration and user-global skill installation. The server used a disposable project, loopback address and dedicated port, without opening a browser.

Test subprocesses excluded model API-key and token environment variables. Search explicitly used semantic:false. No model, embedding provider or GitHub account participated. The reproduction fixture contains commands, identities and raw outcomes.

What the real listener exposed

Initialization created project configuration and a temporary Git repository, with structured output reporting skipped MCP registration. After startup, the web HTML and one referenced frontend asset returned HTTP 200. The document API read back a seeded Markdown file containing Chinese text.

We initialized an MCP session, sent its initialized notification and requested the tool list. The service exposed 19 tools, including search, write, edit, exec and history-related operations. Listing a tool is not evidence that we exercised it.

FlowObserved result
Read seeded UTF-8 document through APIHTTP 200 with expected content
Search an artificial unique termCorrect document and snippet returned
MCP writeApplied/saved response and real Markdown on disk
MCP targeted editApplied response and changed text on disk
Edit absent textError, unsaved/re-read outcome and unchanged file
Write a ../ pathRejected; no corresponding outside file
Create Markdown outside the appBecame searchable through the running service
Stop and start another server processEdited content available through API and search

Serving HTML and an asset verifies frontend distribution, not rich-text interaction. We did not operate the editor, drag components or round-trip complex formatting.

Check saves against actual files

A write containing a heading, artificial marker and Chinese text returned outcome: applied, saved: true and nextStep: none. We then found the Markdown file in the configured content directory and read its bytes. After a targeted edit, we checked the changed disk content again.

An attempt to replace an absent string returned isError: true, with refused, saved: false and nextStep: re-read. The disk file remained unchanged. This gives a caller a concrete recovery direction: reread current content before preparing another edit. Edit implementation

Successful writes also returned preview links and a notice that no browser was attached. We did not automatically follow that notice. Saving content and connecting its preview are separate observable outcomes.

Search and file watching had bounded tests

A lexical query for a seeded unique term returned its expected path and snippet. Creating another Markdown file through the ordinary filesystem made that file discoverable through the running MCP search service. This exercised external-write observation and indexing.

A content-directory .okignore excluded an artificial private/ folder. Searching its unique fake marker returned no result. That demonstrates the selected indexing filter, not authorization or comprehensive confidential-data isolation across all endpoints.

The corpus was small and the terms deliberately distinctive. These were not semantic relevance, Chinese tokenization or large-corpus performance benchmarks. The source describes provider egress when semantic search is enabled; we explicitly excluded that path with semantic:false. Search implementation

What restart recovery establishes

After editing, we stopped the actual server process and started a new one against the same project, then established a new MCP session. The API returned the edited document, and lexical search found it again. The recovered state was a disk document and a usable index, not a model conversation, active task or arbitrary in-memory state.

A write containing ../ was refused, with no test file created outside the content directory. This establishes a boundary for that input through that write entry point; it is not an application-wide security audit. Write implementation

We did not simulate power loss, conflicting simultaneous edits, a full disk or repository corruption. Normal process-restart success should not be generalized into reliability under every failure.

Adoption decision

The release merits a trial when a team already keeps Markdown and wants structured agent writes and retrieval with directly inspectable files. Limit the content directory, choose the search mode and repeat write/recovery checks with representative documents before integrating. The project uses GPL-3.0-or-later.

Desktop installation, model-driven writing, WYSIWYG round-tripping, collaboration, GitHub auto-sync, sharing, remote exposure and other operating systems were outside this review. For another local knowledge workflow, see the Obsidian Second Brain review. For installation-side configuration behavior, see the Claude Code Templates review.