Human-gated agent execution platform
A platform for AI agents that act on outside websites. The first job I gave it is my own job search: it finds postings, drafts applications from verified facts and fills in forms, and every step that cannot be undone waits for my approval.
- Python
- FastAPI
- Celery
- Redis
- Playwright
- PostgreSQL
How it works
- Job feedscompany career pages
- Clean and dedupestored in Postgres
- GradeAI, rules as backup
- Evidenceclaims trace to a profile
- Approvala person says yes
- Final checksstop if unresolved
- Form fillingbrowser, per-site
- Follow-upeach message approved
Decisions that matter
It starts switched off
The default is a disarmed mode. Submitting an application, sending a message and making a paid model call never happen implicitly, and anything that cannot be taken back waits for a person.
Nothing is invented
Every claim in a resume or letter has to trace to a verified profile. When an input is missing, the gap stays visible instead of being filled in.
Behavior changes show up as diffs
Hand-written test cases pin the grading rules and the grading prompt, and every prompt lives in one registry with saved snapshots. A change that shifts behavior fails a check and needs a reviewed diff, so nothing drifts quietly.
Money is reserved before it is spent
A paid model call reserves its cost before it is sent and settles the real usage afterwards, so spend is recorded as it happens.
Limits
Built for one person, by design, and run locally rather than hosted. The spend ledger covers the main model calls but does not yet cap every paid path.
