Osint Agent Skills

OSINT Agent Skills

A knowledge base that turns any autonomous AI agent into a senior OSINT analyst.

[License: MIT] [Status: v1.6.0] [Stars] [Last Commit] [Issues] [Repo Size]

What is this?

osint-agent-skills is not an agent. It is not a script. It is not a SaaS.

It is a structured knowledge base — a curated set of methodologies, tool registries, pivot playbooks, ethics rules, and report templates — that any autonomous AI agent can consume to instantly adopt the operating discipline of a senior open-source intelligence analyst.

If you point Claude Code, Cursor, Ollama, OpenCode, AutoClaw, or any other agent framework at this repository, the agent will:

1. Load system-prompt.md and adopt the OSINT Agent Skills persona.
2. Consult knowledge/methodologies/ to plan its investigation.
3. Use tools/free-tools.yaml and tools/apis.yaml to execute lookups.
4. Follow knowledge/pivot-playbooks/ to chain findings into networks.
5. Generate a report using templates/reports/intelligence-report.md.
6. Respect ethics/legal-frameworks.md throughout — never suggesting illegal techniques, never fabricating findings.

This repository is agent-agnostic. It works the same way regardless of which LLM or agent framework you use.

Why does this exist?

Modern LLMs are powerful OSINT tools — but only when given the right discipline. Without an explicit operating framework, agents hallucinate IPs, skip source citations, blur the line between OSINT and intrusion, and produce reports that look authoritative but cannot be audited.

osint-agent-skills solves this by codifying the operating rules that a senior analyst would otherwise enforce through peer review. The system prompt is brutally explicit about anti-hallucination. The pivot playbooks tell the agent exactly what to do when it finds an email, a domain, a phone number, a cryptocurrency wallet, or a photograph. The ethics framework defines what is in-bounds by jurisdiction. The report templates enforce source citation and confidence labeling.

The result: when you ask an agent that has consumed this knowledge base investigate the domain example.com", you get a structured intelligence report with cited sources, performed pivots, and explicit limitations — not a stream of plausible-sounding prose.

Who is this for?

• Security researchers who want their agent to do OSINT legwork with the rigor they would apply themselves.
• Threat intelligence analysts who want to automate the repetitive collection phase and focus on analysis.
• Journalists investigating disinformation, fraud, or corruption — especially those using Bellingcat-style methodology.
• Due diligence teams who need to background companies and individuals from public sources.
• OSINT educators who want a reference framework for teaching methodology.
• Anyone who has watched an LLM investigate something and produce confidently wrong results.

Quick start

git clone https://github.com/frangelbarrera/osint-agent-skillscd osint-agent-skills

Then point your agent at the repository. Integration guides:

Test prompt after setup:

Investigate the domain example.com using OSINT Agent Skills methodology. Produce a full intelligence report.

The agent should load the system prompt, plan the investigation using the intelligence cycle, run DNS/WHOIS/RDAP/CT-lookups against free tools, follow the domain-to-infrastructure pivot playbook, and deliver a structured report.

Repository structure

osint-agent-skills/system-prompt.md           # The brain agent adopts this personaagent-config.yaml          # Declarative config that agents readREADME.md                  # This fileLICENSE                    # MITCONTRIBUTING.md            # How to contributeCHANGELOG.md               # Version history‚knowledge/                 # The methodology + playbooksmethodologies/         # Intelligence cycle, ATT&CK, Bellingcat, etc.domains/               # Person, domain, IP, company, phone, crypto, ...techniques/            # Dorks, EXIF, Wayback, CT logs, ...pivot-playbooks/       # If you find X, investigate Y"‚tools/                     # Tool registriesfree-tools.yaml        # No API key requiredapis.yaml              # API key requiredmcp-tools.json         # MCP-format for Claude / Cursorcli-tools.yaml         # Local CLI tools (sherlock, holehe, ...)‚templates/                 # Report / plan / evidence / graph templatesreports/               # ICD-203 inspired intelligence reportsinvestigation-plan/    # Scope + plan templatesevidence/              # Evidence logs and chain of custody|   - graphs/                # Investigation graph templates (Mermaid, DOT, JSON)‚ethics/                    # Legal + ethical frameworklegal-frameworks.md    # Per-jurisdiction laws (US, EU, UK, LatAm)jurisdiction-rules.md  # Country-specific rulescode-of-conduct.md     # Investigator code of conductprivacy-guidelines.md  # PII handlinganti-hallucination.md  # Anti-fabrication rules‚case-studies/              # Worked examples of real investigationsintegrations/              # How to wire into each agent frameworkexamples/                  # Walkthroughs of common investigations

What's inside the system prompt?

The system prompt (system-prompt.md, ~3000 words) defines:

• Identity. OSINT Agent Skills — a senior analyst, not a chatbot.
• Core principles. Verify, don't assume. Pivot intelligently. Never hallucinate. Respect legality. Maintain OPSEC. Document everything. Minimize harm.
• Methodology. A five-phase Intelligence Cycle adapted for autonomous OSINT work.
• Anti-hallucination rules. Explicit prohibitions on fabricating identifiers, tool output, dates, or confidence levels.
• Tool usage protocol. How to choose tools, when to ask for paid-quota approval, when to flag techniques for human review.
• Pivot protocol. When to pivot autonomously vs. when to pause for user approval.
• Reporting standard. The default report structure (ICD-203 inspired).
• Ethical boundaries. Hard refusals (stalking, doxxing, pretexting without authorization, deepfake generation, investigating minors outside safeguarding context).
• Output format. No conversational filler. No artificial endings. Source per finding. Limitations section mandatory.

Read it in full: system-prompt.md.

What's inside the pivot playbooks?

The pivot playbooks are the most operationally valuable part of this repository. Each playbook specifies:

• The trigger — what finding activates this playbook.
• The steps — ordered collection actions with tool, command, and expected output.
• The anti-patterns — what NOT to do (e.g., do not assume two identical usernames indicate the same person).
• The output format — how to report the pivot's results.

Current playbooks:

See knowledge/pivot-playbooks/.

Anti-hallucination commitment

This repository was built with one principle above all others: agents that consume this knowledge base do not fabricate findings.

The system prompt explicitly forbids inventing IP addresses, email addresses, usernames, dates, tool outputs, or confidence levels. Each finding in a report must cite a source. Each tool invocation must be real — if a tool failed or returned nothing, that is what gets reported.

If an agent that has consumed this knowledge base fabricates a finding, that is a critical defect and should be reported as an issue. See ethics/anti-hallucination.md for the full rule set.

Output integrity hashing (v1.5.0+)

Every tool response now includes an output_hash field — a sha256 over the canonical JSON form of {tool, endpoint, status, result}. This lets downstream verifiers (audit logs, report reviewers, external scripts) confirm that any data attributed to a tool call was actually produced by that tool call.

The verify_output_integrity tool takes a claimed hash + the response payload subset and returns {valid: true} if the recomputed hash matches, or {valid: false, expected_hash, actual_hash} otherwise. Designed for downstream verifiers — not for in-band LLM self-verification (an LLM that can compute sha256 in its head could forge a matching pair).

Attribution

This repository distills and restructures methodology originally compiled in OSINT-BIBLE by Frangel Raúl Crespo Barrera. OSINT-BIBLE is a curated index of 426+ OSINT resources across 33 sections; osint-agent-skills adapts that material for autonomous agent consumption.

Case studies reference investigations originally published by Bellingcat, Mandiant, CrowdStrike, CISA, and other public-domain threat intelligence sources. Citations appear in each case study file.

Tool descriptions reference the official documentation of each tool. Pricing and rate-limit information was current as of the repository's last update — always verify with the provider before relying on a specific limit.

License

MIT. See LICENSE.

Contributing

See CONTRIBUTING.md. Pull requests that add new pivot playbooks, new free tools, or new jurisdiction rules are especially welcome.

Security

Security issues should be reported privately according to the project's security policy. Please do not open a public issue for a suspected vulnerability.

The latest published npm release is 1.6.0. Versions 1.4.1 and 1.5.0 are deprecated on npm because they predate security and reliability fixes included in 1.6.0.

The project does not claim that every hardening change constitutes a separately exploitable vulnerability. Security advisories will be published only when a specific issue, affected version range, impact, and patched version have been technically confirmed.

Changelog

See CHANGELOG.md.

Agent security model

See docs/agent-security-model.md.