Your AI said no. Get around it anyway.

Your choice of AI

Stop treating one model like the entire engineering department.

Put your own hardware to work

Connect Ollama and choose installed, tool-capable local models for coding. Keep inference on your own hardware when you select an on-device model. Model size, speed, and capability depend on your machine; cloud models and connected tools can still send data off the device.

Bring the AI you're already paying for. BYOK or BYOCLI.

Bring your own API keys, or use supported installed and signed-in coding CLIs such as Claude Code and Codex. Keep the services you already use and add other providers when you need them. Subscriptions and API usage stay with those services.

Retire the model-selection guessing game

Enable automatic model selection, give it a set of models, and describe when each fits. The selector chooses for the first prompt and can reconsider after a failure. Prefer to choose yourself? Pick a provider, model, and supported thinking level, or use /auto to request a fresh automatic choice.

Make models compete

Make your models fight for your trust.

One confident answer is a lousy review process. Use /compare or its alias /review to ask multiple configured models the same question. Choose the contenders and a judge in the desktop dialog, including available local models.

1. Answer independently. Each model proposes its approach.
2. Challenge the answers. Models review the alternatives and can concede to a stronger proposal.
3. Resolve disagreement. If they still disagree, the selected judge ranks the remaining answers. You can inspect the report.

Before a rewrite or an architecture decision, make the models challenge each other's assumptions. Inspect the proposals and review report, then run tests to check the approach you choose.

In the terminal, use --judge with /compare. There isn't a separate /judge command. Comparisons make multiple model calls; provider charges apply.

From idea to something useful

Your app shouldn't die in a chat window.

Open My Apps and Projects, choose Create an app, and describe what you need. Talavine Code creates a project folder, opens its coding chat, and starts the build.

Run or preview a supported app through its configured launch command, return to the build chat, and refine it as you use it. Keep a small tracker, dashboard, or everyday utility within reach instead of losing it in an old conversation.

The files stay in folders you choose. Existing repositories can join the library too. Install the build tools and dependencies your project needs, and configure its launch command.

See how easy it is to open an app and get back to work ↓

Less repeating yourself

Stop being your AI's external hard drive.

The daily reintroduction gets old. Tell Code how you like to work, your preferred tools, or a recurring project convention. Enable personal memory to carry lasting preferences into later sessions, including when you change supported models.

You can inspect and change what it remembers. Edit a memory, hide it from recall, review its history, or forget it. Inferred suggestions wait for your confirmation. Turn memory off for a chat with Temporary (memory off), or disable personal memory in Settings.

Repository memory is separate. Agents read your repository's MEMORY.md and topic files, then propose updates to lasting conventions. Review and apply a proposal before it changes the working tree; committing it is still your decision.

Code also keeps model-response outcome history to inform fallback choices. That history stores request fingerprints and outcomes rather than raw prompts, replies, or credentials.

Personal memory is stored locally. Relevant recalled information is supplied to your selected model, including a cloud provider when you use one. Memory is not the encrypted credential vault.

Work beyond the repository

Give your coding agent hands.

A coding job often ends in a browser: check the page you built, walk through a form, or inspect the result. Code's browser tools can observe pages, click, type, and navigate through a connected Chrome session.

Use your existing Chrome tabs and sign-ins. Enable Chrome's remote debugging connection and approve the connection in Chrome, or choose a separate persistent Code browser profile. You choose which setup to use.

Desktop control reaches native windows too. Code can observe and interact with supported desktop apps through its computer tools. Windows uses the current desktop; the native Mac host requires Screen Recording and Accessibility permissions.

Computer tools follow your permission mode and the operating system's access requirements.

Give it access without repeating secrets

Stop pasting passwords into chats.

Add named credentials in Settings, or import a supported password-manager CSV export. On Windows, the vault encrypts values with a key protected by DPAPI for your Windows user. The native Mac app uses Keychain protection.

Supported tools can use a reference such as {{secret:my-service}}. The host resolves the value when it runs the operation, so you can instruct the agent using a name instead of copying the password into each prompt.

The host also handles the explicit /secret intake command locally, before sending a request to a model. Prefer the credential form for entering a value; you can inspect saved names and account hints without revealing passwords.

Use encrypted storage with a working OS keystore; check the app's storage status. This protects stored credentials and supported secret-reference workflows. It doesn't encrypt all project files, chat history, or personal memory. A password export itself contains sensitive data.

For developers running from source

Make Code work on its own bugs.

Code's developer self-improvement loop (RSI) records defects it encounters and capabilities people request. When enabled and the app is idle, it can take a queued item into an isolated Git worktree, start a repair session, and run configured build and test checks.

A fix earns verified status only after its acceptance checks pass. Follow the session and branch, inspect the evidence, pause the queue, or use the supported merge, rebuild, and relaunch workflow. Automatic merging is a separate configuration choice.

Available in Debug builds run from a source checkout. The public Release installer does not enable this repair loop.

There is also a separate harness-learning workflow in AI overview, with start/pause controls, measured progress, token usage, and rollback. It starts paused for new scopes; it is distinct from the developer repair queue.