Data and privacy
secfoo is local first. This page lists everything it stores, everything that can leave your machine, and how to remove it all.
What secfoo stores
Everything lives in one folder, ~/.secfoo/. Set SECFOO_HOME to put it somewhere else, for example a separate store per client:
export SECFOO_HOME=~/clients/acme/.secfoo
| Path | Holds |
|---|---|
secfoo.db | The SQLite database: projects, assessments, runs, findings, exceptions, acceptances and post-build misses. |
reports/<run-id>/ | report.md, the exact prompt.md the agent received, and the agent's stdout.log and stderr.log. |
attachments/ | Files uploaded to assessments, and the text extracted from vendor documents. |
vendor/ | The Mermaid renderer, after secfoo vendor mermaid. |
config.toml | Your settings and MCP server definitions. |
cloud.toml, memory.toml | Saved Cloud portal and secfoo-memory keys, if you've logged in. Readable only by you (mode 0600). |
What leaves your machine
| When | What is sent | To |
|---|---|---|
| Every run | Whatever the agent reads while reviewing. A CLI agent reads files itself and sends what it reads to its model provider, under that tool's own terms. The api agent puts the target's source files into the prompt. | Your agent's model provider |
| Remote targets | A normal git clone, with your token or SSH key if the repository is private. | GitHub, GitLab or Bitbucket |
| SCA runs, and the SCA dashboard | Each dependency's ecosystem, package name and version, to look up known advisories. Results are cached. | OSV.dev |
Only after secfoo cloud login | Reports and run metadata. See what's sent. | Your Cloud portal |
| Only with a secfoo-memory key | Short, code-free pattern queries: a language, framework, optional CWE and a description of at most 200 characters. | secfoo-memory |
Only when you run secfoo vendor mermaid | A one-time download of the Mermaid renderer. | jsDelivr |
What the agent can do to your code
secfoo asks every agent to review, not change, the target, and limits it where the agent's CLI allows: Claude Code gets a read-only tool grant, Codex runs in its read-only sandbox, and OpenCode has shell, edit and web access denied. Copilot auto-approves tools, so with it read-only behaviour depends on the instructions alone. The agent table has the details for each. For the strongest guarantee, scan a clean checkout or a repository URL (secfoo clones it to a temporary folder and deletes it afterwards).
Scanning code you don't trust
- Claude Code runs with only your user settings and the MCP servers in your secfoo config. A target repository's own
.claudesettings or.mcp.jsonare not loaded. - Reports and diagrams are rendered in the dashboard with raw HTML disabled, and diagrams use Mermaid's strict security level, because their content comes from an agent reading a possibly hostile repository.
- Remote targets are cloned shallowly into a temporary folder. Repository tokens go to Git through a temporary helper and never appear in URLs, arguments, reports or logs.
Backing up and moving
Copy the whole folder while no secfoo run or secfoo serve is running. To move to a new machine, copy it to the same place (or point SECFOO_HOME at it).
Deleting your data
secfoo assessment delete 3 # one case file: its attachments are removed, its runs kept but detached
rm -rf ~/.secfoo # everything: runs, reports, findings, settings and saved keys
There's no command to delete a single run. Uninstalling secfoo doesn't remove ~/.secfoo. If you've synced to the Cloud portal, runs there are managed in the portal.