README.md

# OfferHunt 0.3.5

Job data and tools for personal AI agents. This package exposes sixteen MCP tools through https://www.offerhunt.ai/mcp with OAuth authorization. There is no embedded UI. Conversations may use the user's language; resume parsing currently accepts English PDF/DOCX/text. Ask for the original PDF/DOCX file (a host file reference, or save_resume with no input for a 24-hour upload link); text supports matching but not applications.

Tools: get_profile, save_resume, update_profile, delete_resume, update_preferences, search_jobs, get_recommendations, get_job, rate_job, claim_application, prepare_application_submission, track_application, mark_application_uncertain, mark_application_blocked, release_application_claim, list_my_jobs.

Connection recovery: follow https://www.offerhunt.ai/developers#connection-recovery before implementing or updating a client. The host owns automatic OAuth renewal and secure credential storage. Serialize renewal, persist replacements atomically, and acquire current credentials for each scheduled run. An initial initialize/tools-list 401 may trigger supported renewal/reacquisition, a 10-second delay, and one retry with current credentials. For 503, preserve the connection and respect Retry-After within the run's time budget. Allow one recovery retry in total; no loops, no automatic retries of tool mutations or employer submissions. Persistent rejection needs user attention; a temporary outage does not justify reconnection or disabling a schedule. Use plain-language connection errors. Never collect credentials in chat. Test two expiry/renewal cycles and the actual next-day scheduled run; an immediate setup check proves neither. These instructions do not automatically update previously installed clients or saved tasks.

Agents own scheduling, browser application forms and authorized mailbox reading. OfferHunt returns current jobs, confirmed preferences, private resume URLs and durable application records. An automatic submission requires claim_application before filling the form and a fresh prepare_application_submission check right before submitting. Only confirmed outcomes are recorded as applied with track_application; mark_application_uncertain holds unclear outcomes for verification, mark_application_blocked records obstacles, and release_application_claim frees an attempt only after the user confirms nothing was submitted.

Recommendations wait for answered preferences: when get_recommendations returns needs_preferences, ask its missing_preferences questions together, offer its resume-based suggestions for confirmation, and save the answers with update_preferences (no_preference for "anything is fine"). Users can instead answer on their account page (answer_on_website); it writes the same record, so read get_profile rather than asking again. get_recommendations returns the user's matches a page at a time; the first page holds only strong matches. When more_matches is above zero, say how many more there are and call it with next_cursor when the user asks; later matches score lower. If a cursor returns list_changed, start again without one.

For daily match-only tasks, confirm the user's time and time zone before creating a schedule and report its next run only after host confirmation. Call get_recommendations directly for each run; it reads current matching context, checks pause state, and returns the preferences used. A separate get_profile or application-record read is not required for match-only runs. Application workflows still require fresh profile and application records.

Website: https://www.offerhunt.ai/ · Support: https://www.offerhunt.ai/support · Privacy: https://www.offerhunt.ai/privacy · Terms: https://www.offerhunt.ai/terms