New: every message picks a fast or a slow model

Changelog

What's new in Definable

Every feature, improvement and fix we ship, newest first. The bigger ones get a blog post.

October 2026

Fast and slow models

Each message goes to a fast or a slow model, and either can hand off to the other.

  • NewAdd a fast model beside an agent's main model in Settings. Small talk and lookups go to the fast one, harder work to the main one.
  • NewEither model can hand the conversation to the other mid-run, with a note on what it found.
  • NewTraces open with a Router step: the model that took the message, how sure the router was and why.
  • ImprovedEvery step in a trace is tagged with the model, fast or slow, that wrote it.
  • FixedCitation numbers no longer land on links that contain a number, inside code or on escaped brackets.
From the blogFast and slow models: every message goes to the right one→From the blogSee why a message went to the fast model→

Web search, citations and a stronger browser agent

Agents search and read the web, cite their sources, and do more in Chrome.

  • NewEvery agent can search the web and read pages. It's on by default; turn it off under Settings › Web.
  • NewAnswers that used the web show numbered citations, a source card for each and a Sources list underneath.
  • NewIn Chrome, the agent finds elements by name, runs several steps at once, works inside embedded frames and reads a page's network requests.
  • NewSelect text on any page for Explain, Translate or Ask Definable.
  • ImprovedLong conversations keep working: older turns are summarised once and every model call fits the model's context window.
  • FixedReplies cut off at the model's output limit carry on where they stopped.
From the blogAgents can now search the web and read pages→

Richer answers

Tables, code, diagrams and formulas render everywhere, and answers can be redone and rated.

  • NewAnswers show tables, highlighted code, Mermaid diagrams, math and pictures in the app, the client app and the extension.
  • NewPictures and diagrams open in a viewer.
  • NewRegenerate the latest answer to your message. If it already took actions, you're asked first.
  • NewRate answers with a thumbs up or down; owners see the votes on the run's trace.
  • ImprovedIn the extension: attach files, copy, rate and retry answers, and the agent stays on the tab it started on.

Your own skills and image generation

Write a procedure once as a skill, and models that make images do it in chat.

  • NewMake a workspace skill by uploading a folder, writing it in the editor, importing it from GitHub or asking Builder to draft it.
  • NewModels that can generate images make them right in the chat.
  • NewAgents know when each message was sent, in the sender's time zone, and the sender's language.
  • ImprovedChats with many images open much faster.
From the blogCustom skills: teach agents how your team works→

Live chat turns and onboarding blocks

Watch an agent's work as it happens, and build onboarding from blocks.

  • ImprovedEach turn plays out live and in order: the model's thinking, its words and every tool call as it's written.
  • NewOnboarding is built from blocks: Sign in with Google or an email code, limited to your domains if you like, and ten kinds of question.
  • NewGuardrails: short rules the agent is reminded of before every model call.
  • FixedAn agent that reaches its step limit ends with what it finished and what is left.

A new look and a knowledge drive

A redesigned app, knowledge that browses like a drive, and catalog skills on agents.

  • ImprovedA new design: new type and colours, a new composer and chat history in its own panel.
  • NewKnowledge drive: your bases, files and storage at a glance, each base a folder tree beside the open file.
  • NewAdd catalog skills to an agent. The owner reviews each skill's security audits before approving it.
  • NewChoose whose app account an agent uses: yours, or each customer's own.

An app for your agent's customers

Every agent gets its own page where its customers sign in and chat.

  • NewEach agent has a client app at its own link: customers sign in, finish onboarding, chat and download what the agent made them.
  • NewCustomers can link their WhatsApp, Telegram or Slack chat to the same profile.
  • NewAgents share files through download links that last seven days.
  • NewA Reports tab lists what an agent couldn't do, ready to fix with Builder.

Builder and the Chrome extension

An agent that builds your agents, and agents that work in your browser.

  • NewBuilder: describe what you need and it builds, tries, tests and schedules the agent with you, in chat.
  • NewChrome extension: chat with an agent in Chrome's side panel and let it act on sites you allow.
  • NewOnboarding tab: set the steps people complete before chatting, and see their answers.
  • NewA Run evals first switch beside Publish, on by default, tests a change before it goes live.

Approvals and questions mid-run

Agents can stop and wait for a person, in the app or on a channel.

  • NewAn agent can pause to ask you something, or to get your approval before a tool or app action you chose.
  • NewOn Telegram, Slack and WhatsApp, questions and approvals arrive as messages; reply in words or with YES or NO.
  • NewAn Answer in groups switch for Telegram and Slack.
September 2026

Skills catalog and a live Overview

Browse public skills, watch your workspace live, and use the app on any screen.

  • NewBrowse the public skills catalog with search and sort, and read each skill's files and security audits.
  • NewOverview shows runs, spend, tokens, tool calls and failures, per day and per agent, live.
  • ImprovedEvery screen works from a phone to a wide monitor.

MCP servers, schedules, evals and a computer per chat

Outside tool servers, work that starts itself, tests for agents and a sandbox.

  • NewConnect remote MCP servers, with a token, an API key or OAuth, and give their tools to agents.
  • NewScheduled tasks: run a prompt on a schedule or once, and drag on the calendar to reschedule.
  • NewEvals: write test cases with checks and run them against the real agent.
  • NewPin, rename, archive, branch, delete and share chats through a read-only link.
  • NewAn agent can get its own Linux computer for each chat to run commands and edit files.

Your own model keys and three channels

Run agents on your own provider keys, and answer on Telegram, Slack and WhatsApp.

  • NewConnect your keys for Anthropic, OpenAI, Google Gemini, DeepSeek, Kimi and OpenRouter; each is tested before it's saved.
  • NewTelegram: paste a bot token and the agent answers private chats and messages addressed to it in groups.
  • NewSlack: create the app from a one-click manifest; the agent answers DMs, mentions and threads.
  • NewWhatsApp: link a number with a QR or pairing code for one-to-one chats.

Connected apps and triggers

Agents act in the apps you connect, and start when something happens there.

  • NewConnect your accounts from the apps catalog and link them to agents so they can act in them.
  • NewTriggers: an event in a connected app can start an agent run.