v1.0.4

Security

• Rate limiting reads the client address from the position Railway's edge writes it in X-Forwarded-For and ignores entries further left, which a caller can set to anything. Forging the header no longer risks getting round the limits on other hosts. PROXY_ENTRIES sets how many proxy-added entries to skip.
• The /tighten rate-limit maps are now capped in size, like the public limiter's.
• Runtime dependencies are pinned to exact versions, including mcp 1.30.0, which carries the SDK's published security fixes. Test-only packages have moved out of requirements.txt. Dependabot keeps the pins current.

Methodology: 2026-09 · Full changelog: CHANGELOG.md

v1.0.3

Fixed

• Each tool's title is now also set in annotations.title, where Claude's Connectors Directory looks for it. A test checks every tool has a title and read-only annotations.

Methodology: 2026-09 · Full changelog: CHANGELOG.md

v1.0.2

Changed

• Opening the bare domain in a browser now redirects to the documentation instead of returning Not found. Set DOCS_URL to change where it points.
• The release workflow no longer fails when a release for the tag already exists.

Methodology: 2026-09 · Full changelog: CHANGELOG.md

v1.0.1

Changed

• Hosted endpoint moved to https://footprint.pharmatools.ai/mcp. The previous Railway address keeps working.
• Server instructions and the get_methodology description now describe what the tools return, rather than telling the model how to quote results.
• Privacy section added to the README, covering what the hosted server logs and keeps.

Methodology: 2026-09 · Full changelog: CHANGELOG.md

v1.0.0

First public release.

Added

• MCP server over streamable HTTP at /mcp, with four read-only tools:
  • estimate_ai_footprint: energy and water for chat-style prompts over any number of days, by usage tier and boundary.
  • estimate_session_footprint: an agent or coding session from its token counts, with a 0.3–3× range.
  • compare_water_use: any volume of water in everyday and lifecycle terms, with the boundary caveat.
  • get_methodology: figures, boundaries, sources, limitations, a citation line and why published estimates differ ~100×.
• JSON API mirroring the tools: /api/footprint, /api/session, /api/compare, /api/methodology, with open CORS.
• Hosted endpoint, listed in the official MCP Registry as io.github.nickjlamb/ai-footprint.
• Prompt tightener endpoint (POST /tighten) and its calculator-page embed.
• Per-path CORS and rate limiting: open for the footprint routes, pharmatools.ai-only with daily caps for /tighten.
• /health endpoint, answering GET and HEAD for uptime monitors.
• Test suite pinning outputs to the calculator page (25 prompts a day = 8.5 Wh and 34.8 mL).
• Examples: Python MCP client, JavaScript and curl clients, captured responses.
• Project documentation: contributing guide, security policy, code of conduct, roadmap, citation file, issue and pull request templates.
• CI on Python 3.10 to 3.13 with lint, tests and an end-to-end MCP check. Tagged releases are published automatically from this changelog.

Methodology

• 2026-09: energy per prompt from 2025 vendor disclosures (0.34 Wh simple, 1 Wh mixed, 3 Wh intensive); water at ~1.0 mL/Wh onsite and ~4.1 mL/Wh comprehensive. Typical text prompt ~0.34 mL onsite, ~1.4 mL comprehensive.

Methodology: 2026-09 · Full changelog: CHANGELOG.md