osintkit

Public-data OSINT for one email, username, domain, company, phone number, or Bitcoin address: gathered from free sources, briefed by an LLM acting as a defensive analyst, rendered as a colour-coded report.

osintkit uses only public sources: certificate transparency, public DNS, WHOIS, public GitHub search, free breach indexes, paste archives, and public profile pages. It never logs in on your behalf and never touches a paid data broker.

A full trimmed report, including breach and account sections, is in docs/sample-output.md.

Why

Checking what the public record says about you is tedious: a dozen sites, a dozen formats, no synthesis. osintkit runs those checks concurrently, keeps the raw evidence, and turns it into a briefing with a risk level and concrete hardening advice. Everything is local, every source is named, and every gap is reported instead of hidden.

Intended for personal privacy checks, authorised pre-engagement recon, and quick due diligence on a domain or company. Findings are leads to verify, not facts.

Install

pipx install osintkit
# or
uv tool install osintkit

Every release also carries a wheel with checksums, and pip install -e . works from a clone. Python 3.12+ only.

Quick start

osintkit init                      # pick a provider, paste a key once
osintkit lookup alice@example.com

init stores the provider and key in ~/.config/osintkit/config.toml (mode 600), checks them with a one-token request, and afterwards needs no flags and no exports. osintkit config shows what is stored, masked; osintkit lookup just works.

Environment variables still win over the file, so DEEPSEEK_API_KEY=... osintkit lookup ... overrides the stored provider for one run.

No key? osintkit lookup example.com --offline builds the whole report locally from rule checks, no model involved. --raw gives the collected JSON and --dry-run prints the prompt a model would receive.

Any OpenAI-compatible endpoint works. init --provider ollama needs no key for a local server; --provider, --base-url, and --model cover everything else. The preset table, key variables, and precedence order are in docs/REFERENCE.md.

Optional keys: a Have I Been Pwned subscription (--hibp-key on init, or HIBP_API_KEY) adds per-account breach and paste membership; GITHUB_TOKEN lifts GitHub from 60 to 5000 requests/hour. Both stay local — the key is yours, and every user of the tool brings their own. Everything else needs no credentials.

Commands

Output: default colour report, --json, --markdown (they combine), --raw for collected data, --offline for a local report with no model, --dry-run for the prompt. Exit codes: 0 ok, 1 synthesis or partial failure, 2 bad input.

How it works

target -> type detection -> collectors (parallel, cached, retried)
                                 |
                                 v
                     merge + risk signals (deterministic)
                                 |
                                 v
           LLM briefing -> pydantic validation -> one repair retry
                                 |            (--offline replaces this with local rules)
                                 |
                                 v
              colour report . JSON . Markdown . archive

Collectors degrade to a logged gap, never a dead run; a report is produced even when sources or the model fail.

• Merged, enriched breaches. Records collapse into one row per incident, keeping every source and the earliest date, then pick up dates and exposed-field lists from the public HIBP and XposedOrNot catalogues. Those catalogues are free; only HIBP account checks need a key.
• Email domains follow the address. An email lookup also checks the domain's DNS, mail posture, and WHOIS, marked as derived.
• Info-stealer logs. Hudson Rock reports machines infected by stealers that saved credentials, with dates and hostnames; one infection forces a high risk floor.
• Risk floor. Breach count and credential types set a floor the model may exceed but never undercut.
• Cached everything. 24h default, shorter for volatile sources; repeat prompts cost nothing.
• Validated output. Strict JSON schema, one repair round-trip, then a clean failure.
• Confidence per section. Each section states low, medium, or high, capped by how many sources actually back it, so a single weak source never reads as certain.
• Self-checked claims. Every breach, date, host, and account URL is compared back to the collected data; anything unsupported is called out.

Collectors

Two dozen sources across four target types, all free and login-free: HIBP (key), XposedOrNot, LeakCheck, Gravatar, crt.sh (subdomains, certificates, organisation search), public DNS (A/AAAA/MX/TXT/NS/CNAME plus SPF, DMARC, DKIM, wildcard), passive subdomains, WHOIS, TCP connect probes, GitHub (profile, commits, user search, org repos), psbdmp with a Gist fallback, grep.app, Hacker News, Stack Exchange, and username presence. The full table is in docs/REFERENCE.md or via osintkit collectors.

Breach depth

Breach name lists are free; breach metadata is what makes them useful. Every row is joined against the public HIBP and XposedOrNot catalogues for date, exposed fields, and record count, so the report can say "LinkedIn, 2021-06, emails and password hashes" instead of a bare name. Add a HIBP_API_KEY (about $4/month) to get per-account membership with the same metadata, plus pastes.

Stealer logs matter more than most breaches: the credentials were live on an infected machine, not just leaked years ago. They are queried for every target type, keyless.

Passwords are checked separately with osintkit password: Pwned Passwords covers hundreds of millions of leaked passwords, needs no key, and only the first five SHA-1 characters leave the machine.

Monitoring

osintkit watch --domain example.com --every 1d
osintkit watch --email you@example.com --every 6h --no-notify

The first run prints a baseline. Later runs print nothing unless diff sees movement — a new breach, a new host, a source failing, or a risk change — and optionally raise a desktop notification through notify-send. --cycles bounds the loop, ctrl-c stops it, and every run is archived so you can inspect it afterwards. Pair with cron or a systemd timer for unattended use.

Safety and legality

• Public data only. No logins, no scraping behind auth, no paid brokers.
• Targets are typed and normalised before any request.
• Requests retry on 429/5xx with backoff and honour GitHub quota headers.
• Results are evidence, not ground truth; the report says which sources were unavailable.
• Cache, logs, and reports stay under ~/.osintkit (override with OSINTKIT_HOME). Retention is 30 days by default.

Limitations

WHOIS is often redacted. GitHub search is shallow. Free breach indexes lag. Port probing only sees what your network can reach. Model summaries can overstate weak signals. Read the coverage panel: an empty section may mean "nothing found" or "source skipped".

Development

git clone https://github.com/sjdsjadjajdjsa/osintkit
cd osintkit
uv sync --extra dev
pytest
uvx ruff check src tests

Tests never touch the network: HTTP is mocked with respx, DNS and WHOIS are monkeypatched, and a recorded model reply covers the full-stack path. CI runs on 3.12 and 3.13. See CONTRIBUTING.md for the collector contract and docs/architecture.md for the module map.

Licence

MIT. Security reports: SECURITY.md.

osintkit is independent tooling. It is not affiliated with any source it queries.