Blog
Insights on AI, product updates, and best practices
Fix a Failing Test on Your Own Machine From a Chat
A worked idapt Code run: two permissions, one prompt, and an agent that reproduces a failing test on your laptop, fixes it, and hands back the diff.
Introducing idapt Code
idapt Code turns a normal idapt chat into a coding session on your own machine or an idapt cloud computer, with the model and the tools you already use.
Cancel ChatGPT Plus Without Losing Your History
The switcher's checklist: export your ChatGPT archive, import every conversation into idapt in minutes, and know exactly what maps where before you cancel.
Capability-Matched Cost Cutting: Pay for the Ceiling You Use
Most AI spend goes to flagship models doing non-flagship work. How to match benchmark tier to task tier, verify the swap, and keep quality where it matters.
Conversation Branching: Three Patterns That Beat Starting Over
Edit a message to fork reality, regenerate with a different model, and keep every path: three branching patterns for getting more out of one conversation.
Bring Your Own Keys: BYOK in Practice
How BYOK works in idapt: connect provider API keys, route matching models on your accounts, and mix BYOK with platform billing and local inference.
One Prompt, Five Models: Comparing Side by Side
How to run the same prompt across several models in idapt with parallel tabs, when triangulation beats trusting one answer, and how to pick a keeper.
Computer Use: When the Agent Needs the Screen
idapt agents can now operate a desktop: see the screen, click, type, and verify results, under per-computer consent, an exclusive control lease, and a visible border.
Connect Your Tools to idapt Over MCP
idapt is an MCP server: connect Claude Code, editors, or any MCP client and reach your Drive, agents, chats, and computers through the open protocol.
Cost Controls That Professionals Expect
How idapt makes AI spend legible: pre-send estimates, per-request pricing in the app, run budgets, plan allowances, and a usage log you can audit.
Fan Out Big Jobs With Subagents
How subagent delegation works in idapt: an orchestrator splits work across parallel agents and merges results, with patterns for research, review, and extraction.
Generate Images Side by Side
How to run one image brief across two models in idapt, which model fits which job, and how iteration works when generation lives inside your chats.
Give Your Agent a Memory
How agent memory works in idapt: what agents remember across chats, how to read and correct it, and the difference between memory, context, and files.
How Model Failover Works in idapt
The routing engineering behind 200+ models: cheapest-first provider selection, automatic failover, sticky conversation pinning, and visible attribution.
How the Auto-Router Picks Your Model
What idapt's Auto mode actually does: how it matches a prompt to a model tier, when it escalates, what it costs, and when to pin a model instead.
How We Rank 250+ Models
The methodology behind idapt's rankings: named public benchmarks, a rescaled capability index, live telemetry, and the rule against invented composites.
Your Machine Is a Provider Now: Local Inference
Local inference is GA: pair a computer running Ollama and idapt routes to it as a free provider, with hardware-fit matching and cloud fallback.
What 'Never Trained on Your Data' Means Technically
The engineering behind idapt's privacy stance: no training on user content, database-enforced isolation, a local-inference lane, and blocked AI training crawlers.
New Models in idapt: July 2026
The July model report: OpenAI's GPT-5.6 Sol, Terra, and Luna, xAI's Grok 4.5 flagship, and the catalog crossing 250 models from 25 providers.
From Prompt to Deployed Site on a Cloud Computer
A worked end-to-end run: an idapt agent scaffolds a site, runs it on a cloud computer, fixes what breaks, and exposes it on a public URL.
Real Receipts: How Model Outputs Are Captured
Why idapt's model showcases are captured from real production runs with cost and latency attached, and why staged demos are banned from proof surfaces.
A Research Workflow That Cites Itself
The four-step idapt research loop: sources into Drive, cited search, subagent reading in parallel, and multi-model verification before anything ships.
Run Local Models With Ollama in idapt
The full local-inference guide: pair a machine, pull the right models for your hardware, set prefer-local routing, and verify where each reply ran.
Script Your Workspace: The idapt CLI
The idapt CLI drives your whole workspace from the terminal: chats, files, agents, computers, and generation, with five copy-paste recipes to start.
Build on idapt: SDK Quickstart
The @idapt/sdk in ten minutes: authenticate, chat with any model, read and write Drive files, and run agents from TypeScript.
Teach It Once: Skills Are Generally Available
Skills are now GA for every idapt workspace: reusable instruction bundles your agents discover and apply automatically, written once in a single file.
Tasks Grows Up: Boards, Filters, and My Tasks
idapt Tasks now has drag-and-drop boards, saved filter views, bulk actions, sub-task roll-up, and a cross-list My Tasks page. Here is what shipped.
Team Workspaces: Roles, Sharing, and Shared Context
How teams run idapt: workspaces as the collaboration boundary, member roles, shared agents and files, and the governance controls that keep it sane.
The Autonomy Dial: From Read-Only to Full Control
How idapt bounds what an AI agent may do: four autonomy levels per chat, confirmations on consequential actions, budgets, and complete run traces.
The Model Treadmill
The best AI model changes every few months. Chasing it by switching apps is the treadmill; the exit is infrastructure that outlives any model.
Video Generation Is Live: 16 Models in Your Chats
Generate video in any idapt chat: Veo, Runway, Kling, Luma, and more, from text or an image, billed per second with clips saved to your Drive.
Push-to-Talk AI on Your Desktop: The Voice HUD
The Voice HUD is a global-hotkey overlay for Windows: press, speak, hear the answer from any model, and hand off to the full app when you need more.
Voice Mode, End to End
How voice works in idapt: real-time conversation with any model, barge-in, transcription, spoken replies, and the workflows where voice beats typing.
What Shipped in idapt: July 2026
The July changelog: Skills GA, Tasks boards and My Tasks, faster chat, redesigned cloud computer lifecycle, new comparison surfaces, and more.
Why Real Work Needs an AI Workspace
Chat answers questions; work produces artifacts. Why files, agents, computers, and 200+ models in one workspace beat a standalone chatbot for real output.
Write Your First Skill in Ten Minutes
A worked example of authoring an idapt skill: turn your most-pasted prompt into a single file every agent in the workspace applies automatically.
Your First Automation: An Agent That Runs Monday Mornings
Build a scheduled agent in idapt: a Monday digest that searches, compares against last week, and files a cited report before you sit down.
How to Import Your ChatGPT History Into idapt
A five-minute guide: export your ChatGPT data, upload the ZIP, and get your whole conversation history (branches, tool calls, and images) in idapt.
Use 200+ AI Models with the OpenAI API Format
Point any OpenAI SDK at idapt's compatible endpoint and reach Claude, GPT, Gemini, and 200+ more models through one base URL, one key, one bill.
The catalog grows to 190 models from 23 providers
February 2026: 170+ new models land in idapt, including MiniMax M2.5, GLM-5, Kimi K2.5, and the full GPT-5, Claude 4.x, Grok 4, and Gemini 3 lineups.
Introducing idapt: Your AI Workspace
Meet idapt, the AI workspace where 200+ models, files, agents, cloud computers, and tasks work together. Research, build, delegate, and create.