OpenOSINT is an open-source OSINT (open-source intelligence) framework built natively for AI agents through the Model Context Protocol. An agent like Claude can investigate a domain, a username or an email address by calling OpenOSINT’s tools — instead of a human spending hours across dozens of sources.

The problem

An OSINT investigation means checking many sources by hand, cross-referencing results and keeping track of where each fact came from. It’s slow, repetitive and easy to get wrong.

AI agents can do much of this work — but only if they have good tools. A model with badly designed tools calls the wrong things, wastes context, and produces results nobody can verify.

What I built

  • MCP-native tools for domain, username, email and geographic research, designed to be called by agents rather than wrapped from existing APIs.
  • An entity graph based on FollowTheMoney, with statement-level provenance: every fact records where it came from. Deduplication is non-destructive, with human review.
  • A web UI with chat, graph and globe views, where the map is driven by the agent rather than being a passive dashboard.

The lesson: a critique made it better

A security practitioner pointed out publicly that OSINT agents are noisy: they call too many tools, too aggressively, and leave too many traces on the targets.

He was right. I shipped three changes:

  1. Passive-only mode by default. The agent uses only sources that don’t touch the target.
  2. A cap on tool calls per investigation.
  3. A noise label on every tool, so both the agent and the user know what each call costs.

He tested the release and replied publicly with praise.

The broader lesson, which I now apply to every MCP server: tools for agents must be designed for agents — what they expose, how they’re described and what they return all change how the model behaves.

Security, handled in the open

Running a tool that agents use means treating security as part of the product. I handled two vulnerabilities through coordinated disclosure and published GitHub Security Advisories (GHSA-cqr4-hcfp-m6m4 and GHSA-2wrv-jxxj-r2xc), shipping fixes in released versions.

Results

  • 1.8k stars on GitHub; the first 1,000 came in about 70 days
  • Dozens of releases, shipped continuously by a single maintainer
  • Hosted products built on top of the open-source core at app.openosint.tech

Want something like this for your product?

I build MCP servers that let AI agents use your API, product or data. See packages and pricing →