Career Compass MCP
Career Compass runs entirely on your own computer. It has no server and no account. Your career history and job pipeline are plain YAML files in a directory you choose. Nothing is uploaded, and there is nothing for us to collect, store, sell, or hand over — because your data never reaches us in the first place.
The server makes no network requests of its own. Commands you run yourself with npx,
such as the dashboard command check_setup prints, download the package from npm, and
nothing about you is sent. Details under Updates.
In claude.ai chat (web, desktop and mobile) only the plugin's skills load, as instructions to Claude. The server does not run there, so Career Compass stores nothing and sends nothing. Details under In claude.ai chat.
Career Compass reads and writes the career information you give it:
harvest_evidence to
look at one. It reads that folder's commit history locally (commit authors, dates, and
which files each commit touched) to report what you measurably did there. It writes
nothing, anywhere, and sends nothing.In a single directory on your machine: ~/.career-compass/ by default, or the folder named by the
CAREER_DATA_PATH environment variable when it is set. Files are ordinary YAML you can open, edit, back up,
or delete with any text editor. While a write is in progress the folder also briefly holds a
.write-claim file naming the process doing the writing, so a second Career Compass process
(a dashboard, or a server registered in two clients) refuses rather than overwriting it. It is
deleted when the write finishes, contains no personal data, and is safe to remove by hand.
Each write also leaves a timestamped .bak copy of the
previous version in the same directory. Only the newest 5 .bak files per data file are
kept; older ones are deleted automatically on the next write. Backups you make by hand are
never touched.
The folder also holds .visits.json: when you last used Career Compass, which items on your
daily list you had already been shown, and the last few companies you worked on (a fit
check, a tailored résumé). It is how a return visit can open with what changed while you
were away. Every tool updates it, read-only ones included; it holds no career history,
stays in the folder, is readable only by you, and is never sent anywhere. Deleting it costs
nothing but that note.
We never receive this data. There is no Career Compass account, no cloud sync, no backup service, and no telemetry or analytics of any kind. The server makes no requests to the internet on its own.
Career Compass never checks for updates over the network. check_setup reports the
version you have installed and how a newer one reaches you: the Claude plugin updates
through the Claude directory, and a setup made with npx or install updates when you
run npx -y career-compass-mcp@latest install yourself. Versions up to 2.13.1 could ask
the public npm registry for the latest version number when you asked about updates; that
request is gone.
Career Compass runs nothing on a schedule. Any schedule, such as a morning briefing, is one you create yourself in your own Claude app; it runs locally on your computer and only reads.
Separately, check_setup also checks whether your local dashboard is running by requesting
http://127.0.0.1:<port>/. That is a loopback request to your own machine; it never
reaches the network.
Career Compass is an MCP server, so it answers a client you connect it to — normally Claude Code, or Cowork on your computer. When you ask Claude to tailor a resume or prep an interview, the relevant parts of your career data are passed to that client, and from there to the model provider under their privacy policy, not this one:
That is the only path by which your data leaves your machine, it happens because you asked for it, and it is governed by the terms of whichever client and model provider you chose. Career Compass itself sends none of your data anywhere.
The bundled local dashboard (career-compass-mcp dashboard) serves pages from
127.0.0.1 on your own machine, renders them with no external assets, and makes no
network calls. Starting it means running it, usually with npx, which downloads the
package from the npm registry.
If you start the dashboard with --ask-claude, its buttons run Claude Code on your
computer. Claude Code sends the Career KB content it reads to Anthropic, under your own
Claude account and billing, like any other Claude Code session. It is off unless you pass
the flag, and read-only unless you start it with --ask-claude-writes.
The Career Compass plugin is also offered in claude.ai chat on the web, desktop and mobile. There, only its skills load: plain-text instructions that tell Claude how to do a fit check, tailor a résumé, or prep an interview from what you paste into the conversation. The MCP server above does not run in chat, so:
Job postings, emails, and documents you paste in are third-party content. Career Compass wraps them in a clearly-marked, nonce-delimited block before they reach the model, so text inside them is presented as quoted evidence rather than as instructions. This reduces the risk that a malicious posting can direct the model to act against you, but no such measure is absolute. Treat pasted content from unknown sources with the same care you would apply anywhere else.
Your files stay on your disk until you delete them. The one exception is backups: only the
newest 5 .bak files per data file are kept, and older ones are deleted automatically.
There is no retention period on our side because we hold nothing. To remove everything, delete your CAREER_DATA_PATH
directory (including the .bak files, .visits.json and any leftover .write-claim) and uninstall the package.
None. We do not share, sell, rent, or disclose your data, because we never receive it. There are no advertisers, analytics providers, data brokers, or subprocessors involved.
Career Compass is a job-search tool intended for adults in the workforce and is not directed at children under 13.
Changes are published in this file in the public repository, with the date at the top updated. The version history is visible in git.