DEV Community

Ajnas N B
Ajnas N B

Posted on

Three open-source building blocks for AI agents: memory, safer tools, and web access

AI agents are becoming capable enough to work across whole projects, but three practical gaps keep appearing: they forget why decisions were made, tool approvals can drift away from the action that was reviewed, and web access is often either too limited or too broad.

I have been building three open-source projects around those gaps.

Qarinah - evidence-linked project memory

Qarinah keeps project decisions, code structure, tool outcomes, approvals, and sources in a local record. For each task it compiles a compact, cited context pack instead of replaying the entire project history.

Website: https://qarinah.io

GitHub: https://github.com/AjnasNB/qarinah

Maqam - policy and exact approval for tool calls

Maqam checks policy before registered operations, binds approval to the exact run, tool, and input, and records traces and evidence after execution. It is designed to sit around agent workflows rather than replace the framework or model you already use.

Website: https://maqamagent.com

GitHub: https://github.com/AjnasNB/maqam

Cockroach Crawler - web access for AI agents

Cockroach Crawler is a JavaScript toolkit for static and browser crawling, structured extraction, PDFs, screenshots, site maps, source adapters, MCP, Docker, and agent integrations. It gives agents useful web reach while keeping network authority and resource limits with the host application.

Website: https://cockroachcrawler.com

GitHub: https://github.com/AjnasNB/cockroach-crawler

Why they fit together

Qarinah remembers what happened and why. Maqam decides whether a registered action may run. Cockroach Crawler collects web evidence the agent can use. Each project is independently usable, but together they form a practical memory, action, and research layer for agent systems.

All three are open source. I am looking for developers who will run them in real workflows, report failures, challenge the boundaries, improve adapters and documentation, and contribute code.

If you build with AI agents, which problem is most painful in your current stack: memory, tool safety, or web access?

Top comments (0)