What's new in Runtype

Runtype MCP endpoints now serve the handshake-free 2026-07-28 MCP revision alongside 2025-11-25 on the same URLs, with cache hints, deterministic tool ordering, and RFC 9207 issuer validation.

Reasoning support is now read from the model catalog, Claude 4.6+ uses adaptive thinking, model refusals route through your fallback chain, Mixlayer Qwen models accept images, and new accounts get Gemini 3.6 Flash.

A personal-shopper chat agent for any Shopify storefront, driving Shopify's built-in WebMCP tools (search, cart, checkout) through a Persona widget you add with one script tag.

Parallel AI-judge scoring, abort-proof sync runs, judges that see seeded conversations and check results, per-case graders, hand-written tool mocks, and batch support for conversation-seeded cases.

A secret reference picker with inline creation, "Get key" links to each provider's console, a hardened value mask, and product setup screens that show every key your product actually references.

Custom tools created via API, MCP, or SDK now appear in the agent editor's tool picker, and code mode is configurable from the editor sidebar.

npx -y @runtypelabs/cli@latest onboard connects Claude Code, Cursor, Codex, Gemini CLI, OpenCode, or VS Code to Runtype, and agents can now complete signup with just an email and a 6-digit code.

create_surface and update_surface now accept inbound and outbound credential objects, so headless setup can configure webhook signing secrets and messaging-provider keys without the REST API.

A webhook surface with an empty event list now accepts every event instead of rejecting payloads with "Event not allowed".

Logs now break down each execution's context into system prompt, tools, history, tool results, and message, per model call and per agent iteration, with actionable hints when it gets heavy.

runtype install-mcp detects your coding agent, installs the matching Runtype skill, configures hosted MCP, and starts OAuth when the client supports it.

Global Runtype CLI installs can now stage a health-checked stable release for the next invocation, with manual check, update, and rollback controls.

Detached subagent runs return a durable run ID immediately, keep working after the parent request ends, and report completion or approval needs back to the parent conversation.

Build branching flows with named paths and fallbacks, insert variables and encrypted secrets in place, and control tool-call and stream-step failure behavior.

Define a JSON Schema for a record type, validate writes in warn or enforce mode, and use the same collection workflow through REST, SDK, MCP, and CLI.

Paste your Slack App Credentials once, click Add to Slack, and let Runtype complete the OAuth exchange instead of copying a bot token by hand.

New product and surface setup reports give coding agents structured pending steps, recommended actions, and headless alternatives.

Embedded agents can search catalogs of up to 200 page tools and refresh the available toolset after navigation, so larger cross-page tasks fit in one run.

Connect MCP servers that require an API key plus tenant, project, base-URL, or routing headers, with encrypted secret references for credentials.

Set a monthly dollar cap, see credit and billed spend at a glance, and understand where usage and prompt-cache savings come from.

Prompts now cache automatically on supported models. Repeated context bills at the provider's cached rate, and the Usage page shows your hit rate, cached vs uncached tokens, and estimated savings.

AI graders now judge the whole run with evidence-backed verdicts, route to a judge from a different model family by default, and can be split into focused checks or replaced with your own judge flow.

Capture any real agent execution as a test case: fork at a point in the conversation, replay with recorded tool results, and let user feedback propose regression cases automatically.

Human-in-the-loop approvals get more precise and more dependable: gate only the tools you choose, enforce the same policy on every execution path, and let approvals wait durably for a decision.

A new redaction capability keeps personal data like emails, phone numbers, and card numbers out of your traces, logs, and telemetry. Turn it on with one setting.

New accounts start on Claude Sonnet 5, the model catalog includes the Claude 5 family, and a hardened migration path moves existing flows and agents off older Sonnet versions.

Two new flow steps make record reads unambiguous: get-record always returns one record, list-records always returns an array. The old retrieve-record step is deprecated but keeps working.

Human approvals, wait-until polls, and async crawls now run on durable Cloudflare Workflows. Long pauses survive deploys and worker restarts, and approvals can wait up to 7 days.

Define evals next to your flows and agents with defineEval, converge them with evals.ensure, and gate CI with the runtype eval run command. Trace graders assert on tool calls, step order, and cost.

Email agents preserve thread context, reply safely to the right audience, and can author sanitized, inline-styled HTML when Markdown is not enough.

We rebuilt the dashboard from the ground up for a faster, more responsive experience. Pages load instantly, navigation is snappy, and there are no more full-page reloads.

Large conversation and Record attachments move to tenant-scoped storage transparently, while supported terminals render their images inline.

Marathon CLI sessions are now branchable and steerable. Browse and fork session history with /tree and /fork, redirect the agent mid-turn, and never lose context that was trimmed from the window.

Keep agent and flow definitions in your repo and converge the platform to them at deploy time. defineAgent and defineFlow plus a hash-first ensure() verb give you reviewed, versioned, drift-checked configs.

Connect Granola as a first-class integration: agents and flows can read meeting notes, transcripts, and folders, and a ready-made template reacts to new meetings on a schedule.

Deduplicate provider retries, combine rapid message bursts, and pause automation so a person can take over one conversation or an entire messaging surface.

Let agents retain useful facts, preferences, and decisions across conversations, with configurable per-agent, per-user, or per-end-user memory scopes.

Persona chat widgets can now discover and call tools your page registers on document.modelContext, so an embedded agent can search your catalog, read page state, and drive your UI.

The Python SDK now includes a Fern-generated low-level API core from Runtype's OpenAPI spec, while preserving FlowBuilder, streaming, and local-tool ergonomics on top.

Agents can use timezone-aware time tools, account for gaps between messages, and schedule themselves to run again later with set_reminder.

Runtype A2A Surfaces now speak the A2A Protocol v1.0 wire format, with 1.0 Agent Cards, PascalCase JSON-RPC methods, typed StreamResponse events, and dual-version import support.

The TypeScript SDK 3.0 and 4.0 releases now derive core response and streaming contracts from Runtype's committed OpenAPI spec, removing stale hand-written entity shapes.

Six composable agent skills for building, managing, embedding, and operating AI products on Runtype — installable in one command across 35+ AI coding agents.

Scheduled agent runs now support inline message sets (including named variants), structured batch inputs, or record-backed runs—so cron and one-shot schedules can drive agents the same ways you use them interactively.

Run `runtype auth export-key` to print the API key stored by the CLI—handy for scripts, CI secrets, and quick env wiring without opening the dashboard.

Transform-data steps now validate JavaScript/TypeScript syntax at save and dispatch time—catching typos and parse errors before sandbox execution.

E-commerce API and agent templates so you can prototype integrations without standing up a full backend.

An auto-generated, version-tied SDK reference—including FlowBuilder chains, dispatch, and eval APIs—ships in-repo and through MCP-style resources so coding agents stay aligned with @runtypelabs/sdk.

Per-request model overrides now apply correctly when executing virtual flows, so dispatch-time model settings match what you pass from the SDK or API.

A dedicated management API for listing, creating, and managing conversations across all channels — widgets, surfaces, and apps.

The dashboard logs UI supports live tail streaming and sharable URL-synced filters, so you and your team can follow runs and debug without leaving the browser.

Flows can invoke agents inline with the new execute-agent step type, combining deterministic routing with autonomous agent turns in one workflow.

Use /copy-trimmed or the session menu to copy the same marathon state as JSON with large code and tool payloads shortened—easier to paste into an LLM without blowing the context window.

Built-in tools for creating, reading, updating, and deleting records—plus retrieve-record query mode for many rows in one step—give agents structured access to your data without custom glue.

Slack connects through a guided, step-by-step setup wizard instead of a raw OAuth-only path—clearer milestones, fewer dead ends, and more reliable onboarding.

Runtype MCP adds scheduling, flow management, and get_me for identity—so external AI clients can automate more of the platform.

Persona 3.0 replaces flat widget theme props with structured design tokens, createTheme(), built-in presets, and component-level theming for embedded chat experiences.

Persona is now open source. Themeable, pluggable, zero-framework chat widget with Shadow DOM isolation, SSE streaming, agent loops, and voice input.

Persona can now stream rich artifacts over SSE, letting agents emit markdown documents and interactive components alongside normal chat responses.

The model catalog has been refreshed with upstream config updates and Qwen 3.5 migration work, keeping defaults and availability current.

Messaging surfaces now get trait-driven behavior and flexible inbound and outbound formatting, making delivery smarter across email, Slack, SMS, Telegram, and more.

Runtype MCP now ships both platform context and modern transport support, with resources and prompts for guidance and Streamable HTTP for live progress.

The product generator now shows architecture previews and interactive clarifying pills, turning product planning into a more guided conversation.

Marathon CLI now combines an Ink TUI with context compaction, session search, tool input streaming, workflow plugins, and upgrade enforcement.

Flows can now run a first-class Cloudflare Browser Rendering crawl step with progress streaming, async pause/resume, and page-level billing.

Leaked Runtype API keys found by GitHub Secret Scanning can now be revoked automatically through a dedicated webhook pipeline.

Bring your own OpenAI-compatible model endpoints, including local or hosted deployments, and route custom models through Runtype with BYOK controls.

Managed Orthogonal platform tools let flows and agents call paid third-party APIs like Apollo and LinkUp with built-in spend tracking.

A2A is out of beta with Agent-to-Agent protocol 0.3 alignment, stable agent card URLs, and consistent artifacts and status across the task stream.

Runtype now supports server-side ElevenLabs speech-to-text and text-to-speech tools, enabling voice input and spoken agent responses.

Connect OAuth2-protected MCP servers with standard auth code and PKCE flows, encrypted credential storage, and automatic token refresh during execution.

Agents can now run multi-turn reasoning loops across API, chat, and MCP with configurable turn limits, reflection, cost budgets, execution history, and telemetry.

The new Code Mode MCP endpoint exposes just search and execute, cutting MCP context overhead dramatically for coding assistants.

Sensitive JSONB fields are now encrypted at rest, protecting stored credentials, tokens, and webhook secrets with transparent seal and unseal handling.

Transform-data steps now execute in sandboxed Cloudflare Workers (V8 isolates) for better security and async/await support.

Real-time visibility into flow and agent execution with the new Events panel in Persona.

The built-in vector-search tool adds semantic retrieval to agents and flows across pgvector, Weaviate, and Vectorize-backed knowledge bases.

Per-customer execution limits with rate limiting, daily caps, monthly caps, and spend limits. Configurable per plan.

Define tools inline with dispatch requests: external HTTP endpoints, custom sandboxed code, or nested flow execution.

Support for A2A as a surface so agents can talk to agents: tasks, SSE streaming, agent cards, and skill-based routing—now in beta.

Client tokens let you embed Persona in browsers without exposing API keys, with origin restrictions and generated JavaScript and React snippets.