What is Kensaku?
Kensaku is an open-source desktop search app for Windows, macOS and Linux that indexes file names, the text inside PDFs, Office files, EPUBs, email and more than 50 code and data formats, text read from screenshots by OCR, and a small embedding model for search by meaning. Its kensaku-cli binary runs an MCP server over stdio that searches the same local index, and can also serve it over loopback HTTP.
The repository, from the GitHub user yjid, was created on 24 October 2026 and had about 1,645 stars and 107 forks by 11 November. It is MIT-licensed and written in Rust with a Tauri interface. Version 1.0.17, released on 9 November 2026, is a security release that fixed five privately reported issues, three of them rated high.
How to install Kensaku
kensaku-cli mcp (kensaku-cli ships with each GitHub release)Is Kensaku worth it?
Kensaku gives an agent something most file servers do not: search by what a file says rather than what it is called, across documents, code and the text inside screenshots, from an index a desktop app already keeps current. The MCP tools are search, regex_search, find_symbol for tree-sitter definitions in Rust, Python, JavaScript and TypeScript, read_snippet, health_check and index_state. All are read-only. A reindex tool exists but is hidden and off unless you set mcp_allow_reindex in the config, and it never runs over HTTP.
The responses are built for a context window. Search snippets are capped at 100 characters, pages default to 10 results and stop at 50, and read_snippet returns at most 1,024 tokens centred on the first match rather than a whole file. According to the project's benchmarks, a name search across 100,000 files averages 18 ms on Windows 11 with default threads, and the metadata for those files takes 12.2 MB. The query language (type:pdf, modified:7d, re:"...") passes straight through the search tool.
The catch is age. The repository was created on 24 October 2026, and 1.0.17 on 9 November fixed five advisories, including agent root restrictions that could be bypassed through search scopes and find_symbol, read_snippet reading excluded files that were never indexed, and a Linux updater that accepted an attacker-owned staging directory. Fixing them quickly and publishing advisories is a good sign; finding three high-severity holes in the agent boundary in the first fortnight says the boundary is still being worked out. Release builds are unsigned on macOS and Windows.
The prompt-injection exposure is the index itself. Whatever text sits in a watched folder, including downloaded PDFs and other people's screenshots, can come back to the agent in a snippet. The docs say file content is returned as data, never as instructions, but that is a label on the output, not a filter the model is bound by. Credential files such as .pem and .key, folders like .ssh and .aws, and .env files are denied by default, and agent_scoped_roots narrows what an agent can see to chosen folders; anyone pointing an agent at a home directory should set it.
Where it earns its place
A backend developer chasing a deploy failure from last week
Find the log line and the screenshot of the error without remembering file names
Ask the agent to search for the error text with modified:7d; Kensaku returns matches from log files and OCR'd screenshots with paths, and read_snippet pulls the surrounding lines. Free; the index is already built by the desktop app.
A consultant with years of client PDFs and decks on a laptop
Pull every proposal that mentions a particular pricing model
Set agent_scoped_roots to the clients folder, then have the agent run a content search filtered to type:pdf and summarise the hits from snippets. Nothing leaves the machine except what the agent's own model provider sees.
An engineer new to a large mixed-language repository
Locate where a function or struct is defined across several projects
Use find_symbol with a language filter to get the path, line range and signature, then open only those files. Faster than a full grep across a home directory, and limited to 200 parsed files per call.
Before you commit
- Setup
- Download the app bundle for your platform from GitHub releases (unsigned, so expect a SmartScreen or Gatekeeper prompt), pick folders and exclusions on first run, then point an MCP client at kensaku-cli mcp. Settings, Agents can write the entry for Claude Desktop, Claude Code and Cursor with a .bak backup.
- Integrations
- Six visible tools: search, regex_search, find_symbol, read_snippet, health_check and index_state, plus an opt-in reindex. Stdio by default; kensaku-cli serve adds MCP Streamable HTTP on loopback port 8765 with a bearer token.
- Data and privacy
- Reads the index of folders you choose, which holds file names and extracted text; tools cannot change files. No telemetry; outbound traffic is an update check and model or OCR pack downloads. The HTTP mode opens a loopback port with a token stored at <config>/http.token. Indexed text is untrusted input to the agent. Five advisories (three high) were fixed in 1.0.17.
- Cost at scale
- Free. The cost is local: a first index of about 101,000 mixed files took 184 s at 183 MB peak in the project's Linux reference run.
- Limits
- About 16 open issues. Exhaustive mode caps at 10,000 results per walk; find_symbol covers four languages; JPEG 2000, fax and JBIG2 scans are not decoded; visual search is a beta download outside results. Windows gets the most real-world testing.
What’s good
- Searches document text, OCR and meaning, not just names
- Read-only tools; reindex off by default
- Token-budgeted snippets and pagination
- No account, telemetry or cloud search
What’s not
- Five security advisories fixed in its first weeks
- Unsigned builds on macOS and Windows
- Indexed text can carry prompt injection to the agent
- Visual search not yet part of results
Best for
Developers and knowledge workers who want Claude or Cursor to find files on their own machine by content, and who will scope the agent to chosen folders.
Kensaku alternatives
Kensaku: common questions
- Is Kensaku worth it?
- A fast way to let an agent find what is on your own disk by content, from a project barely three weeks old whose first security release has only just landed. We rate it 3.7 out of 5.
- How much does Kensaku cost?
- Free and open source (MIT). No paid tier, account or cloud service.
- What are the downsides of Kensaku?
- Five security advisories fixed in its first weeks; unsigned builds on macOS and Windows; indexed text can carry prompt injection to the agent; visual search not yet part of results.
- Who is Kensaku best for?
- Developers and knowledge workers who want Claude or Cursor to find files on their own machine by content, and who will scope the agent to chosen folders.
- How much does Kensaku cost for a team or in production?
- Free. The cost is local: a first index of about 101,000 mixed files took 184 s at 183 MB peak in the project's Linux reference run.
- Is Kensaku safe for company data?
- Reads the index of folders you choose, which holds file names and extracted text; tools cannot change files. No telemetry; outbound traffic is an update check and model or OCR pack downloads. The HTTP mode opens a loopback port with a token stored at <config>/http.token. Indexed text is untrusted input to the agent. Five advisories (three high) were fixed in 1.0.17.
- What are the best Kensaku alternatives?
- The closest alternatives we have reviewed are Filesystem (reference server), Zvec-Grep (zg) and Paperless-NGX MCP Server.
Answers checked Nov 11, 2026.