Built this: AgentForge

On 7 October 2026 I started AgentForge. It has 67 commits so far, more than any of my other repos.

How it works

From the README, three CrewAI agents run in sequence:

agent	job
Researcher	searches the web and reads pages with real tools
Analyst	separates evidence from interpretation and lists gaps
Reviewer	audits the analysis, records failed checks, writes the final report

The design rules the README spells out:

• Sources are never written by the model. The sources list is recorded by the web_search and fetch_page tools, and each source is labelled “page read” or “search result only”.
• Progress is real. The timeline is driven by events from the running crew.
• Failures are reported as failures. Errors, timeouts and missing sources end the job as failed with a reason.
• Async jobs: POST returns 202 with a job id and the UI polls for results.
• fetch_page only accepts public http(s) addresses, as a basic SSRF guard.

Stack: React + Vite + TypeScript frontend, Python + FastAPI + CrewAI backend.

Status, as the README states it

Built and tested, but not yet run against a live LLM or deployed as a hosted app. The screenshots come from a stubbed test backend.