Apply Agent — AI skill for Claude Code

apply-agent

An LLM agent that keeps track of job applications by reading an email inbox, **read-only**. It classifies replies (rejection, interview invitation, request for information, offer, other), extracts company / role / date / summary, and keeps a table of applications and their current status.

[](https://github.com/jbsanchezr/apply-agent/actions/workflows/ci.yml)  [](LICENSE)

**Stack:** LangGraph · Claude or a local model via Ollama · FastAPI · SQLAlchemy · Prometheus · Grafana · Langfuse · Docker · GitHub Actions

Highlights

• Runs with zero credentials. Out of the box it reads bundled synthetic emails and uses a keyword baseline, so docker compose up works on any machine. A free local LLM (Ollama) or Claude is one environment variable away.
• Measured, not assumed. An evaluation harness runs the whole agent over 26 labelled emails and reports per-class precision/recall, confidence intervals and a confusion matrix. The local qwen3:8b model goes from the baseline's 73% accuracy to 100% (with an honest caveat, below).
• Read-only by design. The Gmail integration asks only for gmail.readonly, rejects broader tokens, and exposes a read-only interface. Email bodies are never stored or logged.
• Contained against prompt injection. The model gets only two tools, both bound to the email being processed, so a malicious email can at worst get itself misclassified.
• Observable. Prometheus metrics (tokens, cost, latency, outcomes), a 19-panel Grafana dashboard that is tested against the code, and optional Langfuse traces.
• Reproducible CI for free. LLM responses are recorded once and replayed, so CI re-runs the LLM evaluation on every push without a GPU or API key.
• Documented trade-offs. 45 desi