Skip to main content

Privacy & security

devcoach is local-first and single-user by design. Everything it does happens on the machine where your agent runs.

Everything stays on your machine​

  • All your data lives in one SQLite file: ~/.devcoach/coaching.db (plus an optional ~/.devcoach/learning-state.md notebook). Nothing else, nowhere else.
  • devcoach makes no network calls of its own and sends no telemetry. Your knowledge map, lessons, feedback, and project/git context never leave your computer.
  • It runs as a local stdio process started by your agent — there's no server to expose, no account, and no login.

Why it can't be a hosted / remote connector​

claude.ai web and other "remote MCP" connectors require a multi-tenant HTTP/OAuth server. devcoach deliberately writes to your local home directory, so it cannot run as a remote connector — that local-only constraint is the whole point. On claude.ai you can still use the skill copy, which provides the coaching behaviour without any data storage.

Backups are plain files​

devcoach backup produces an unencrypted zip (settings, knowledge map, lessons, and the notebook). Lessons can include snippets of your project context, so treat backups like any other sensitive working file and store them somewhere you trust. See Backup, export & import.

Sharing is explicit​

Sharing a lesson is the only way a lesson leaves your machine, and it only happens when you ask for it:

  • Only the lesson travels by default — title, summary, body, topic, categories and level. Project, branch, commit and task context are included only with an explicit Include where it happened / --with-context / include_context; a local folder path is never exported, and a repository name only when it points at a remote host (GitHub, GitLab, Bitbucket).
  • The share link is server-less — the lesson is encoded in the URL fragment, which browsers never send to the site; the docs page decodes it locally. devcoach still makes no network calls of its own: the only outbound request is devcoach import <url> fetching a URL you pasted.
  • Your name is opt-in — proposed from share_name or git user.name, editable, and "anonymous" is always available.
  • Nothing is saved on the receiving side until you click Add to my lessons (or run the import).

What devcoach reads​

During automatic onboarding devcoach reads a small set of metadata sources, locally and read-only, to suggest topics — never the content of what you typed:

  • the ~/.claude.json projects map — project paths only;
  • each project's manifests and lockfiles, via a depth-limited directory walk;
  • ~/.claude/history.jsonl — only the project and timestamp fields of each line, to rank projects by recent activity; the prompt text on those lines is never read;
  • at most 1200 characters of each project's auto-memory MEMORY.md.

Only the 15 most recently active projects are scanned, and any failure degrades to an empty scan. Git metadata — project, repository, branch, commit — is auto-detected from your working directory and stored only in your local database, purely to give lessons useful context. devcoach itself makes no network calls; the only subprocesses it runs are git (for that metadata) and, for the plugin / extension bootstrap, a one-time npm install of the pinned package.

Automatic (Deep) onboarding is the one explicit exception. It's an opt-in tier you choose at setup time, clearly labeled as such. If you pick it, the skill first calls preview_deep_scan — a metadata-only pre-check that lists which projects fall inside the chosen date window and lets you narrow it — and only then spawns a separate subagent that reads real conversation text from your local Claude Code session history (~/.claude/projects/<project>/*.jsonl, at most the 5 most recent sessions per project). That subagent returns only synthesized topics, groups, and notebook notes — the raw transcripts never enter your main conversation. Nothing leaves your machine except as part of your own conversation with the model, same as anything else in that session. Standard Automatic and Guided onboarding never read this content — only Deep does, and only when you choose it.