We help product teams create, test, and ship AI experiences. Beyond demos - real products, in production, serving customers. Make It Real with Runtype.
Persona.js V5 preview is here 👀
The VanillaJS agent UI library: open source, no framework dependencies, as unique as your product.
Works with AI SDK, OpenAI Agents, Claude Agent SDK, LangGraph, Mastra, Flue, Eve or any SSE stream.
What’s missing? 👇
https://t.co/30fSOm1omv
The worst agent failures don't always throw errors.
We're building Runtype Trace Signals with @Cloudflare's Clairic + Clef Flash to find frustration and unanswered questions.
The goal: analyze every trace, at scale, without breaking the bank.
https://t.co/zLIUQhzoRj
Stop asking a chat model to read your invoices. Put @Cloudflare's new Clairic to work in Runtype flows: extract typed fields, branch on confidence, and save the results. No generated tokens. Now with an Extract Fields step.
https://t.co/rYTszP9jpJ
Putting an AI agent behind an @mention in Slack doesn't solve the noise problem. It just moves the routing work onto your team.
Someone still has to notice a message matters, remember which bot handles it, and tag it in.
Here's how we're using Jev and small decision models to move beyond the @mention: https://t.co/Pbg27vhsY1
Work across Runtype organizations without logging out. The CLI keeps multiple accounts, links each project to the right one, and uses your OS keychain where available. A linked project will not silently run under another org.
https://t.co/it4UN1Rb0t
Use decision models for structured questions about text or JSON. Discover models, publish a question-set version, and run it through the API, SDK, or MCP, with per-question refusals preserved. Dashboard authoring is a separate rollout.
https://t.co/YLiGyObQUM
Let a mid-run message find the right place. Opt-in auto routing can steer a durable agent with a correction or queue a separate task for later. If the classifier cannot decide, the message queues instead of being dropped.
https://t.co/a6j6Qr6vz9
Connect custom models that speak the OpenAI Responses API. Choose Responses or Chat Completions in your inference connection, including from the dashboard. Existing connections keep their current format.
https://t.co/38fA4iNSfa
Ask why an agent run behaved that way, then inspect the evidence. Explain this run separates observations, possible causes, and next steps, with citations into the trace. You choose whether to act on the explanation.
https://t.co/rzB8Fio1Ti
Keep more of your AI product in source control. Pull deployed definitions, preview changes, and fail CI on drift. Expanded CLI support and Python/Ruby ensure helpers cover flows, tools, surfaces, products, skills, and evals.
https://t.co/I6fNTp2Ax4
Test the same agent as different users and tenants. Eval cases can now carry their own identity, with suite defaults when you need them. Check personalized behavior and identity-aware tools before your users do.
https://t.co/VdZcDQceLT
Let your APIs verify who an agent is acting for. Runtype can send a signed, short-lived user and tenant identity token to your tools. The model cannot choose the identity, and your service keeps control of authorization.
https://t.co/uBt6zEMkgB
Bring your existing agent into Runtype without rebuilding it. Connect a compatible A2A endpoint, or use a local CLI listener during development. Keep your framework; use a public endpoint for channels and scheduled runs.
https://t.co/h6HiQk2DTQ
Require zero data retention for model inference across your workspace. Runtype uses eligible routes and refuses unsupported ones. An admin controls the policy in Settings > Models; your Runtype log and record retention stays separate.
https://t.co/JajzouNRiW
Take a versioned agent to your own infrastructure. Export its definition, bound skills, and prebuilt Runtype runtime in one bundle, with a manifest and file hashes. Available with Enterprise self-hosting.
https://t.co/rkPJT0s2H5
See what your agent actually sent to the model. Explore context block by block, compare requests across a run, and inspect captured text alongside token estimates and cache coverage. Your logging and privacy settings still apply.
https://t.co/xpC6fgFV7X
Run Claude Code or Codex inside your agent's Linux sandbox. Both come preinstalled, with managed inference and no provider keys to copy into the container. Your organization's model-key policy still applies.
https://t.co/b6T2lEImkQ
@Runtypical@Cloudflare@stripe Coming soon: a free template of the eval harness we built to continuously test our own AI onboarding, so any team can use it.
Follow along so you don't miss it.
Last night our CPO @Runtypical joined @Cloudflare and @stripe at Cloudflare's SF HQ to talk about selling to agents. The room was full of tech leaders, startup founders, and retailers. 🧵
@Runtypical@Cloudflare@stripe Hugo Buret from Stripe showed Stripe Projects buying products and services end-to-end in Muse. He also made the case that agents call today's business and pricing models into question: "Agents get a budget, not a subscription."
@Runtypical@Cloudflare@stripe@mgamache from Cloudflare showed agent-first trials. She also covered the tools Cloudflare gives app builders and site owners to keep the internet a level playing field as automation pushes traffic up 10x. Some of that traffic is good, some is harmful.