A 6-phase, 56-task OSINT workflow with a Python task runner, source grading, hypothesis analysis, templates and about 15 scripts.
Unusually thorough procedure and tooling, but it is dual-use (username enumeration, archiving), installs packages system-wide, has few users and has not been updated since May.
Sleuth turns Claude into a structured investigator. A task_runner.py script is the gate: the agent runs next to read the current task, does the work, then runs done, which checks a checklist before advancing. The 56 tasks sit in six phases: operational direction, intelligence collection, collation and entity resolution, chronological and relational processing, hypothesis and reasoning, and a final report. The agent keeps an investigation notebook as its working memory.
Three analytic standards are mandatory in the instructions. Every claim gets an Admiralty 6x6 grade for source reliability and credibility. Conclusions come from Analysis of Competing Hypotheses, where the surviving hypothesis has the least evidence against it. Probability wording must map to the ICD 203 seven-tier scale. Collection is justified through a PLAN check on proportionality and legality before it begins.
The skill folder is large: a 20 KB SKILL.md, about 65 reference files, 23 templates and 18 scripts. Scripts handle source grading, entity resolution, network graphs, timelines, sanctions screening, domain and WHOIS lookups, geolocation, Companies House queries, username enumeration, media archiving with yt-dlp, and report generation. Several call public APIs, and it can install Python packages on demand. There are no tests.
A 20 KB SKILL.md with about 65 reference files, 23 templates and 18 purpose-built scripts and a gating task runner, though no tests or evals.
Counted from the repository file tree. How we score depth.
This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.
assess tool with the current profile state and an evidence block whenever subject information is synthesised, so investigation content goes to a third-party server (server.smithery.ai/elbpr, elb-pr-contact.workers.dev) with no per-use opt-in.skills/claude-sleuth/SKILL.md:49get_technique, get_thinking_toolkit, get_framework) to load methodologies and scoring criteria from the server and apply them, so those instructions are not in the repo.skills/claude-sleuth/SKILL.md:47An OSINT and investigation skill with 15+ Python tools that query public registries and search for usernames across many sites. It automatically pip-installs packages into the system Python, tells the agent to send evidence about the subject to third-party hosted MCP servers and to load methods from them, and ships a self-hostable Cloudflare Worker database whose /mcp endpoint has no authentication. The README also lists a different GitHub owner for the clone URL than the repo it sits under.
pip install <package> --break-system-packages for any missing module (requests, aiohttp, lxml, pandas, yt-dlp, playwright, weasyprint and others) into the system Python.skills/claude-sleuth/scripts/task_runner.py:106wrangler CLI.skills/claude-sleuth/scripts/username_enum.py:38The repository (127 files) on Oct 29, 2026. Not a full review: we read 8 of 21 code files; it relies on code outside the repository: Hosted MCP servers named in server/mcp.json and SKILL.md (server.smithery.ai/elbpr/thinking-toolkit, .../pigeon-superstition-superposition, and the elb-pr-contact.workers.dev deployments); Third-party tools and PyPI packages the scripts install or call (maigret, sherlock, yt-dlp, gallery-dl, playwright, wrangler, requests, aiohttp, lxml, pandas and others); The other 13 skills/claude-sleuth/scripts modules (chronological_matrix, corporate_intel, domain_intel, entity_resolver, financial_analysis, geolocation, network_graph, report_generator, sanctions_screen, source_grader, template_builder, config, __init__) were searched for network, process and file calls rather than read line by line.
NVIDIA SkillSpector static scan: 48 pattern hits (Supply Chain 38, Data Exfiltration 6, Dangerous Code Execution 2, Agent Snooping 1, Excessive Agency 1). The 38 LOW hits are unpinned dependencies in pyproject.toml; 'mcp.json' is the README telling you to add the MCP server URLs to .claude/mcp.json; the external-transmission hits are public data API endpoints (SEC, GLEIF, Open-Meteo, Open-Elevation, USAspending); the two subprocess hits are the automatic pip installs.
A review, not a certification. Skills change after we check them — read the files before installing.
MIT — commercial use allowed.
Not really. The task runner gives a resumable, gated sequence, but each task expects human-set subject, scope and legal justification, and the checklist items are reviewed interactively.
A single research skill with an 8-phase pipeline, parallel search, source scoring, claim and citation checks, and report output in Markdown, HTML and PDF.
Plugin packages of skills for journalists and researchers (verification, FOIA, AP style) plus dev, newsroom-ops and security toolkits for Claude Code and Codex.
Sleuth is a Claude skill that runs an OSINT investigation as a 56-task intelligence-cycle workflow. A task runner gates each step, and the skill supplies templates, reference files and scripts for source grading, entity resolution, timelines and reporting.
On Claude Desktop, mobile or web, upload the skills/claude-sleuth folder as a skill. For the CLI, clone the repo, run pip install ".[all]" and use python3 scripts/task_runner.py next to start.
We found no hidden credential reads or obfuscated code. It does make network requests to public lookup services, can install packages with pip --break-system-packages, and includes tools like username enumeration that can be misused, so run it in a virtual environment and keep to lawful, proportionate investigations.
Sleuth is a process-first workflow with grading and hypothesis standards. elementalsouls/Claude-OSINT is a set of reconnaissance skills focused on breadth of capabilities. Choose Sleuth for auditable analysis and the other for recon tooling.