A quick tour of what's landed on calypr.co since the MVP push started — five weeks, five themes.
Week 1 — Measurement
Product analytics wired end to end, because you can't improve what you can't see. Every run, template load, and code export is now an event — including the ones that matter most to us: when you open the Code view and take the Python with you.
Week 2 — Metering & durability
- Real usage accounting: every run records its tokens and cost, per node and per model.
- Durable conversations: playground threads now survive server restarts — checkpoints moved from memory to Postgres.
- A platform-wide spend kill-switch, so a runaway loop can't run up a bill.
Week 3 — Share-to-test links
The first way to put your agent in front of someone else. Mint an unguessable link and anyone can chat with your agent — logged out, streamed, and capped per link. The graph itself never leaves the server: visitors get the conversation, not your design.
https://www.calypr.co/s/<token> → chat with the agent, nothing elseMarkdown rendering landed in the shared chat too, so replies with lists and code read the way they should.
Week 4 — Fail gracefully
Partner-readiness polish: friendly, leak-safe error messages (no more raw stack traces in chat), toasts on failed saves and runs, and error boundaries so a crash degrades to a card instead of a blank page. A cyclic-graph validator now catches infinite loops before any model call burns tokens.
Week 5 — The reverse round-trip begins
The existential feature. Calypr has always compiled canvas → code; now the reverse parser
has landed: parse_python(code) recovers the graph — topology, state channels, and layout
— from generated (even hand-edited) Python. Every shipped template round-trips in CI.
This is the foundation for editing code and applying it back to the canvas, coming in the next few weeks. No ceiling, in both directions.
Want to try any of this? Open the canvas — the fake model needs no API key.