The first five weeks were about making Calypr trustworthy — metering, durability, share links, honest errors. Since then the work has been about making agents useful: giving them senses, a voice, and a way to reach the tools you already use.
Agents that see, draw, and speak
Three new blocks, each one drag-and-drop on the canvas:
- Image — turn a prompt into a picture. A
stylefield pins the look, so an agent can be permanently an anime illustrator without you writing a word of prompt engineering downstream. - Voice — text to speech, with an
instructionsfield that steers tone and pacing. The Translate & speak template chains it after a translator agent: one run gives you the Chinese transcript and the spoken audio. - Upload — attach an image and let a vision agent read it. Ships with two starters, Label & receipt reader and Generate alt text, that are just this block plus a prompt.
Generated images and audio render inline in the playground and on share links — a player, not a URL.
Web search that actually searches
The Tools block now runs Tavily for real. Add your key in Settings and your agent searches the live web; without one it says so plainly rather than inventing results. Two more providers landed alongside it:
- Unsplash — real photos by search term. The Image Finder template previews them inline.
- Generic HTTP — point a tool at any JSON API you like.
MCP: one plug shape for everything else
The bigger change is MCP support. MCP is a standard socket for tools: a vendor publishes an MCP server for their product, and any app that speaks the protocol can plug in and immediately ask "what can you do?" — no per-vendor integration code.
Your Tools block can now point at any MCP server over HTTPS. Paste a URL (and a token if it needs one), and the tools that server offers become tools your agent can call. It's encrypted at rest and resolved server-side at run time — the secret never touches your graph or the generated code.
Notion, in two clicks
The first connector built on top of that: connect your Notion account from Settings → Connectors. It's a normal OAuth flow — you approve it on Notion's side, we store the token encrypted, and your agent gets all 24 Notion tools: search pages, read them, create them.
The Notion assistant template is a working starting point — ask it about your workspace and it goes and looks.
One agent, several tools
Agents can now be wired to more than one Tools block at once, and each call is dispatched to the block that owns it. So a single agent can search the web and read your Notion — a research assistant that checks your own notes before it checks the internet.
This is subtler than it sounds, and getting it wrong is quiet: the earlier version bound every tool correctly but could only route calls to one block, so the agent would insist it had no access to something it was, in fact, holding. Fixed, and pinned with tests that fail if it regresses.
Bring your own keys
Settings → API Keys lets you paste your own OpenAI, Anthropic or Tavily key. When present, your workspace's key is used instead of ours — shown masked once saved, encrypted at rest, never returned by any API. It's also how the frontier models (Opus 4.8, Kimi K3) run: those are BYO-key only.
Quieter, but worth knowing
The canvas validator got stricter about graphs that would look fine and then do nothing — a Tools block wired somewhere that can't bind it, a branching step whose paths aren't labelled. These now fail loudly on the canvas, before a run spends anything, and name the block to fix. A test suite wires every pair of block types together to keep it that way.
Everything above is live at calypr.co. If you build something with the new connectors, tell us what broke — that's what the next changelog is made of.