Clearfront 2.7.3

Added

• --version flag. The CLI could not report its own version: --version exited 2 with the usage block, and -v is --verbose. Both -V and --version now print the installed version and exit.

Changed

• Every runtime dependency now caps the next major, and CI gained two guards against the class of failure below: a weekly scheduled run, since a dependency can break the published package with no commit behind it and nothing was running while nobody pushed, and tests/test_entrypoints.py, which imports every shipped entry point. The MCP break was invisible to a cold-install check that only exercised the CLI.

Fixed

• The MCP server was broken on every fresh install. The mcp dependency was declared as mcp>=1.28.1 with no upper bound. mcp 2.0.0 landed on 2026-07-28 and changed the Server API, so clearfront/mcp_server.py failed at import with AttributeError: 'Server' object has no attribute 'list_tools'. Any pip install clearfront from that date resolved to the new major and could not start the MCP server; the CLI was unaffected. The constraint is now mcp>=1.28.1,<2.
• An empty Bright Data response is a failure, not a finding. A suspended or unpaid Bright Data account answers POST /request with HTTP 200 and a zero-length body. Every caller read that as content, so one dead backend produced three different misleading errors across scrape_url, search_footprint, search_dorks_live and search_paste, and the analyst reported the outage as an absence of evidence. Empty bodies now raise, and the error names the account state from GET /status. The SERP tools also carry the underlying error into their aggregate message instead of blaming your credentials.

Clearfront 2.7.2

Added

• Agent skill at skills/clearfront-osint/, following the Agent Skills open standard, so Claude Code, Cursor, Gemini CLI, Copilot, Goose, OpenCode and the other skills-compatible clients can drive Clearfront directly. Install with npx skills add scottmartinanderson/clearfront. The skill covers choosing an approach for a given request, the direct subcommands, how to read the confidence and severity ratings, and the authorized-use rules. It declares its own behavior up front (outbound calls are the point, no credential or browser-profile access, no telemetry, and the ollama provider keeps a run entirely local), and it verifies Clearfront is present before running anything, since a virtualenv install reads as "command not found" rather than as a missing tool.
• Plugin manifest at .claude-plugin/plugin.json, so Clearfront can be installed as a plugin rather than a bare skill.
• MCP Registry entry. The README now carries the mcp-name: sh.clearfront/clearfront token, claiming the DNS-verified sh.clearfront namespace on the official registry rather than a personal-account namespace, and server.json describes the package for registry-fed directories.

Changed

• README documents both integration paths and the line between them: the agent skill (instructions for driving the tool) versus the MCP server (direct access to all 30 tools).

Clearfront 2.7.1

Metadata release. Names Clearfront OSINT in the package description, keywords, MCP server and README so the PyPI mirrors and MCP directories carry the full name.

No functional changes.

Clearfront 2.7.0

Sweep effort in the REPL and CLI

The Faster / Balanced / Deeper control that shipped in the web console now works for the terminal analyst too.

• Set it with the --effort flag: clearfront --effort faster
• Or change it live in the REPL with the effort command (bare effort shows the current level)
• Each level caps the tool rounds (4 / 8 / 12) and shapes the enrichment instruction, so lighter sweeps read deliberate, not truncated
• Deeper stays the default with its behavior unchanged; OIS_MAX_TOOL_ROUNDS still sets the Deeper ceiling
• Works across all three providers (Anthropic, Ollama, OpenAI-compatible)

Renamed to "effort"

The sweep-depth control is now called effort across the web console (chip tooltip and labels), the REPL, and the CLI. The three levels (Faster, Balanced, Deeper) and their descriptions are unchanged.

See the CHANGELOG for details.

Clearfront 2.6.0

Sweep depth control

The web console header now has a sweep-depth control that sets how far each investigation fans out:

• Faster — fewer tool rounds, a quick pass over priority sources, light on your API budget.
• Balanced — a moderate sweep of the main sources.
• Deeper — the full exhaustive map. This is the default, and its behavior is unchanged.

The menu explains each level and its time and cost tradeoff on hover and on keyboard focus, and your choice persists across sessions. Depth changes collection breadth only, never the analyst's reasoning or output quality, so anyone running on their own API key can dial down token spend on lighter sweeps without touching the deep default. OIS_MAX_TOOL_ROUNDS still sets the Deeper ceiling and drives the CLI.

Full changelog: https://github.com/scottmartinanderson/clearfront/blob/main/CHANGELOG.md

Clearfront 2.5.0

Open-source AI intelligence on your digital footprint. Clearfront runs 30 OSINT tools, an AI security analyst, and a 3D evidence graph over a name, username, email, phone, domain, or IP, then hands back one calibrated intelligence report. It runs locally, uses your own API keys, and sends your investigations nowhere.

The fastest way to understand it is to point it at yourself and see what the internet already knows.

Highlights in 2.5.0

• 30 collection tools across usernames, email and breach exposure, infrastructure, certificates, and on-chain and infostealer signals, with new keyless additions (search_crt, search_crypto, search_hudsonrock) that need no API key.
• One analyst, everywhere. The REPL, CLI, and multi-target agent now produce the same calibrated ICD-203 intelligence report that used to exist only in the web console, from a shared prompt and tool registry.
• Interactive 3D evidence graph with click-to-detail and in-place pivoting, exportable to GraphML, JSON, and Mermaid (opens in Gephi, yEd, and Maltego).
• Cross-tool corroboration. When independent tools surface the same profile, the analyst flags the agreement so genuine overlap reads as higher confidence, without ever merging or altering raw tool output.
• Report export in the web console: copy, Markdown, and PDF.
• Per-run tool cache so a repeated lookup within one investigation is served from memory instead of the network.

Run it four ways

• Interactive AI REPL: clearfront
• One-shot CLI investigation
• Local web console
• MCP server, so any MCP client can drive the tools

Install

git clone https://github.com/scottmartinanderson/clearfront
cd clearfront
pip install -e .

See the README for the full tool table and quickstart, and clearfront.sh for more. Authorized use only: run it on yourself, or on targets you have permission to investigate.

Full changelog: CHANGELOG.md