Auto-Fix Agent
How it works, and what it can and cannot touch
The five steps
- 1Connect GitHubOne click. We only get access to the repo you pick.
- 2We scan your siteCrawler access, structured data, AI index, page titles.
- 3The agent writes the fixesReal code for your framework, not a to-do list.
- 4It opens pull requestsOne per area, with an explanation of every change.
- 5You review and mergeNothing touches your site until you approve. Then we re-check.
Your code stays yours
- Only the repos you pickA GitHub App you install on selected repositories. Permissions: read/write contents and pull requests, read metadata — nothing else, no other repos, no secrets.
- Reads at most 50 files, 100 KBOnly the files that decide AI visibility — robots.txt, llms.txt, head/layout templates, sitemap and config. Never .env, never your application code at large.
- Stores none of itRepo contents live in memory for the minutes the job runs, then are gone. We keep the fix list, PR numbers and a hash of your live pages to detect deploys — not your code.
- Pull requests only, on a citeflow/* branchThe agent never writes to your default branch. Nothing changes on your site until you review and merge.
- Short-lived tokensEach job gets a GitHub installation token that expires within an hour and is never written to disk.
- Revoke in one clickUninstall the CiteFlow app from GitHub → Settings → Applications. Access ends instantly; we're notified and pause the job.
Questions
Which repositories can CiteFlow see?
Only the ones you tick during the GitHub App install. You can change that list any time from GitHub → Settings → Applications → CiteFlow.
What exactly does the agent read?
An allow-list of GEO-relevant paths — robots.txt, llms.txt, head and layout templates, sitemap and framework config — capped at 50 files and 100 KB. It does not crawl your source tree and never reads .env files.
Is my code stored on your servers?
No. File contents are held in memory for the minutes a job runs and discarded. The database keeps the list of fixes, PR numbers and links, and a hash of your live homepage/robots.txt/llms.txt used to detect that a merge has deployed.
Is my code sent to an AI model?
The files the agent read are sent to Anthropic (Claude) to write the fix, in the same request, and are not retained for training. Every job is capped at $5 of model usage.
Can the agent break my site?
It cannot deploy anything. Changes land on a citeflow/* branch as a pull request with an explanation of every file. Your default branch changes only when you merge, and you can close a PR without merging.
What happens after I merge?
We wait for your deployment, re-scan the site, and show the new score next to the old one on the job page. You also get 5 manual re-checks per job.
How do I revoke access?
Uninstall the app from GitHub → Settings → Applications, or suspend it. Access stops immediately and we receive the event.