The complete list of changes below.
Added
- MCP Connections — Connect Claude or ChatGPT to Framework and your assistant works with your real project data: search a drawing set, check quantities, draft an RFI, upload a drawing.
- REST API — Behind the connectors sits a REST API with scoped, revocable keys for automation and integrations.
- Annual Billing — Every plan now offers a monthly/yearly choice at checkout.
- The Complete CSI MasterFormat Catalog — Cost code pickers in estimates, measurements, RFIs, and submittals now draw on the full MasterFormat 2020 hierarchy — 6,192 sections across 34 divisions.
Changed
- One Overage Rate — Token overage rate is has been reduced by over 50%. Tokens included in each plan have been increased by approximately 30%.
- Much Longer Conversations — Chat now makes real use of the model’s 1M-token window.
- More Reliable Web Search — Chat web search moved to a Perplexity-backed tool after the previous provider search was found intermittently killing runs mid-answer.
- AI Generated Plans — The AI will now generate work plans prior to commencing work.
Fixed
- Encrypted PDFs — Password-protected drawing sets failed in two different ways: some were wrongly rejected as “corrupted,” others reported success while sections of the document were unreadable to the AI. Both paths now decrypt and process cleanly.
- Stuck Conversations — various fixes to improve chat reliability to resolve mid stream pausing.
- Uploads Stuck at “Compressing” — Certain legacy PDFs failed compression and hung forever with no visible error. Compression now falls back through gentler levels — or proceeds with the original file — and any processing failure shows up in the failed-uploads list instead of hanging silently.
- Plain-Language Errors — “This document is too large for automatic AI analysis” now arrives in those words instead of a raw provider error, and a declined card at checkout reads as a declined card rather than a server fault.
- Ongoing Production Triage — A daily review of production errors ran throughout both months; the fixes above are the user-facing results, alongside dozens of smaller corrections.
Security
- One Permission Layer Everywhere — Chat tools, the REST API, and MCP connections all route through a single capability layer that checks project membership and role on every call — the app, your assistant, and the API can’t disagree about what you’re allowed to do.
- MCP Hardening — Only tokens issued through the OAuth connection flow can open the MCP door (an ordinary browser session token can’t), every connection is rate limited, and all connector activity is logged and monitorable.
