> ## Documentation Index > Fetch the complete documentation index at: https://ade-app.dev/docs/llms.txt > Use this file to discover all available pages before exploring further. # AI providers > Connect coding agents with your own key or subscription, and reach 160+ more models through OpenCode — fully managed by ADE. ADE runs your chats, CTO tasks, and workers through **coding agents** you already know — Claude Code, Codex, Cursor, and Droid — plus **OpenCode**, a managed layer that opens up subscriptions, provider keys, and 160+ more models behind one roof. Connect at least one before starting agent work. Settings → Providers is organized into two groups: the coding agents, and everything OpenCode unlocks. ## Coding agents Each coding agent uses the login or key you already have. ADE detects your existing CLI credentials automatically, so most of these light up green without any extra setup. You do not install the agents yourself: ADE downloads its own tested copy of each one for the machine that runs it, the first time that machine uses it, and keeps it at that version. Your own CLI's sign-in is reused. | Agent | Connect with | | - | - | | **Claude Code** | Your `claude` login — a Claude Pro/Max subscription or `ANTHROPIC_API_KEY`. | | **Codex CLI** | Your ChatGPT sign-in — a Plus/Pro subscription or `OPENAI_API_KEY`. | | **Cursor** | A Cursor API key (`CURSOR_API_KEY`). ADE verifies it and lists Cursor's models. | | **Droid** | Your Factory login or `FACTORY_API_KEY`. | | **Antigravity** | Google's Antigravity agent: sign in from its page with a Google account, Gemini Enterprise, a Gemini API key, or Agent Platform. | > [accordion] Claude Code > > Sign in with a Claude subscription or paste an Anthropic API key. ADE runs Claude through the Claude Agent SDK and its own copy of Claude Code, so the `claude` CLI does not need to be installed. If sign-in is required, the card offers a one-click login. See [Claude Code](https://ade-app.dev/docs/ai-tools/claude-code.md) for lane context and permission modes. > [accordion] Codex CLI > > Sign in with your ChatGPT account or provide an API key. Codex runs as a CLI-backed session and reaches OpenAI models. > [accordion] Cursor > > Add a Cursor API key (or set `CURSOR_API_KEY` in the environment ADE launches from). ADE verifies it with the Cursor SDK and discovers your available models. See [Cursor](https://ade-app.dev/docs/ai-tools/cursor.md). > [accordion] Droid > > Sign in with Factory or set `FACTORY_API_KEY`. Droid runs as an agent provider and supports AGI mode for larger autonomous tasks. > [accordion] Antigravity > > ADE downloads Google's Antigravity agent for the machine that runs the chat and signs it in from **Settings → Providers → Antigravity**. Pick a method (Google account, Gemini Enterprise, Gemini API key, or Agent Platform) and choose **Sign in**. The login is the agent's own and is separate from the Antigravity app and the `agy` CLI. > > Signing in to another machine from this one: open the Google link anywhere, then paste the full address the last page ends on (it starts with `http://127.0.0.1` and will not load) into the return-address field and choose **Continue**. It holds a one-time sign-in code, so paste it only there. Gemini Enterprise's license page opens only in a browser on the machine that runs Antigravity. > > Models come from your Antigravity account. Antigravity cannot rewind or edit earlier turns, and it has no read-only Plan mode in ADE; use its own `/plan` command. > [accordion] GitHub Copilot > > Copilot chats and Copilot terminals run the same Copilot version. Copilot's own slash commands work as they do in its CLI: the `/` menu lists the commands and skills Copilot reports, `/context`, `/usage` and `/env` answer without a model request, and `/review`, `/plan` and `/compact` run through Copilot. Commands ADE handles itself, such as `/model` and `/permissions`, point you to ADE's own control. A failed request shows the HTTP status, the error code, and GitHub's request IDs. ## Custom providers: any harness on a plan you have A **custom provider** runs one harness (the agent program, such as Claude Code or Codex) on a model from somewhere else: another provider's sign-in, a provider signed in inside OpenCode, or a stored API key. Run Claude Code on your ChatGPT plan, or Codex on your Claude plan, with the sign-ins you already have. There is no extra login. Open **Settings › Providers › Custom**. With nothing saved yet it offers one-click starters built from your sign-ins. The editor shows a live preview with the harness, the model, and what it uses (the plan and email, or the key's name), then numbered steps. Saved custom providers appear as cards, are saved to your ADE account so every machine has them, and show up as the first tab of every model picker. ADE lends its local proxy only the short-lived access token of each Claude and Codex account, plus your Grok, Kimi, Devin and Muse logins. When a token is about to expire, ADE has that provider's own CLI renew it, so the renewal token never leaves the CLI and your normal chats stay signed in. Copilot, Pi and Kimi take a custom model through their own provider config. Cursor stays on its own sign-in, and Antigravity cannot be used as a source. > [!NOTE] > > On macOS, ADE does not read a Codex login kept in the Keychain, because reading it would prompt you. If Codex stores its login there, the launch explains how to switch Codex to file storage. ## OpenCode — universal model access OpenCode is the managed layer that reaches everything else. ADE downloads and runs it for you, keeps its credentials in an isolated ADE-managed store (never your own `~/.local/share/opencode`), and refreshes its model catalog on its own. One group, four ways in: **1. Subscriptions** Connect a subscription you already pay for — currently GitHub Copilot, ChatGPT Plus/Pro, GitLab, Poe, and DigitalOcean. Click **Connect**, finish the sign-in in your browser, and the row flips to Connected when the provider reports back. The available list comes straight from OpenCode, so it grows over time. **Kimi for Coding** connects here too, with a membership key instead of a browser sign-in. > [!NOTE] > > Claude Pro/Max cannot be used through OpenCode — Anthropic's terms of service don't allow it. Connect Claude through the **Claude Code** agent above instead. **2. API provider keys** Paste a key for any first-class provider — Anthropic, OpenAI, Google AI, Mistral, DeepSeek, xAI, Groq, Together AI, OpenRouter, or **Moonshot AI** (`MOONSHOT_API_KEY`). Use **Verify all** to check every stored key at once. Keys are held in local encrypted storage. **3. More providers** Beyond the first-class rows, OpenCode reaches **160+ more providers**. Search the chip cloud, then paste a key against any provider by its ID. The catalog is cached between sessions, so it renders the moment Settings opens. **4. Local model servers** Point ADE at **LM Studio** or **Ollama** running locally. ADE auto-detects the OpenAI-compatible endpoint, lists the models you've loaded, and lets you set a default per runtime. ### Catalog freshness The group header shows when the model catalog (from models.dev) last synced. ADE refreshes it on launch and every six hours; hit **refresh** in the header to pull the latest metadata on demand. ### Advanced — custom providers & model slugs Open **Advanced** to add a custom OpenAI- or Anthropic-compatible provider (name, base URL, key, and model slugs) or to register extra model slugs by `providerId/modelId`. Both flow straight into the managed OpenCode config. ## The models you can reach Between the coding agents and OpenCode's routing, you get models from every major lab. Pick whichever fits the task. ![Anthropic Claude](https://mintcdn.com/ade-ac1c6011/lmMpn49wx_3bhn_c/media/logos/models/anthropic.svg?fit=max\&auto=format\&n=lmMpn49wx_3bhn_c\&q=85\&s=9961fc27cb76594234ea2b46281dff91) ![OpenAI](https://mintcdn.com/ade-ac1c6011/h8tcu_q5iZ_LGW1L/media/logos/models/openai.svg?fit=max\&auto=format\&n=h8tcu_q5iZ_LGW1L\&q=85\&s=53e6bae29a99cda478387e0a331e7a21) ![Google Gemini](https://mintcdn.com/ade-ac1c6011/lmMpn49wx_3bhn_c/media/logos/models/gemini-color.svg?fit=max\&auto=format\&n=lmMpn49wx_3bhn_c\&q=85\&s=a9da7bbd2f36258e6c193b2722420e3b) ![DeepSeek](https://mintcdn.com/ade-ac1c6011/lmMpn49wx_3bhn_c/media/logos/models/deepseek-color.svg?fit=max\&auto=format\&n=lmMpn49wx_3bhn_c\&q=85\&s=fc647782fe197d79f49c9626f07f5714) ![Qwen](https://mintcdn.com/ade-ac1c6011/h8tcu_q5iZ_LGW1L/media/logos/models/qwen-color.svg?fit=max\&auto=format\&n=h8tcu_q5iZ_LGW1L\&q=85\&s=351ee0ecc17dc337c8b7ccbc60d61000) ![Mistral](https://mintcdn.com/ade-ac1c6011/h8tcu_q5iZ_LGW1L/media/logos/models/mistral-color.svg?fit=max\&auto=format\&n=h8tcu_q5iZ_LGW1L\&q=85\&s=a8fedcaaa546ec5ccdbb4fadbf91f4d6) ![Meta Llama](https://mintcdn.com/ade-ac1c6011/lmMpn49wx_3bhn_c/media/logos/models/meta-color.svg?fit=max\&auto=format\&n=lmMpn49wx_3bhn_c\&q=85\&s=8a15123c0793e6b89e338099059dfb6e) ![Grok](https://mintcdn.com/ade-ac1c6011/lmMpn49wx_3bhn_c/media/logos/models/grok.svg?fit=max\&auto=format\&n=lmMpn49wx_3bhn_c\&q=85\&s=4f49c6d8a130dd613d5fd0778aefe4f1) ![Ollama](https://mintcdn.com/ade-ac1c6011/h8tcu_q5iZ_LGW1L/media/logos/models/ollama.svg?fit=max\&auto=format\&n=h8tcu_q5iZ_LGW1L\&q=85\&s=fbcc53dc68f2759726ac5fe24148d97f) ![Perplexity](https://mintcdn.com/ade-ac1c6011/h8tcu_q5iZ_LGW1L/media/logos/models/perplexity-color.svg?fit=max\&auto=format\&n=h8tcu_q5iZ_LGW1L\&q=85\&s=dcdc8199c8debf954648bdf42c3290e7) ![Cohere](https://mintcdn.com/ade-ac1c6011/lmMpn49wx_3bhn_c/media/logos/models/cohere-color.svg?fit=max\&auto=format\&n=lmMpn49wx_3bhn_c\&q=85\&s=850e8493194001bd5e1b8986ecca6bf0) A model that still needs a key or sign-in shows dimmed in the model picker with a set-up shortcut that jumps straight back to this page. ## Permission presets Providers expose different native permission models. ADE maps them into a small set of practical choices. | Preset style | Use it when | | - | - | | Plan / read-only | You want the agent to inspect and propose first. | | Default / edit | You trust the lane and want normal coding flow. | | Full access | You are in a trusted environment and want minimal prompts. | | Custom | You want the provider's own config to decide. | Set stricter modes for unfamiliar repositories. The full mapping lives in [Permissions](https://ade-app.dev/docs/configuration/permissions.md). ## Cost and usage Running several agents in parallel can spend quickly. Use cheaper models for routine follow-up and stronger models for planning, architecture, and hard debugging. Usage is counted for your whole ADE account. Tokens, spend, requests and active days on the new-chat card, the home page, **Settings › Usage**, and each provider's page add up every machine you are signed in on. An account is shown by its email. Quota belongs to the login, not the machine. The top-bar **Usage** popup has one panel per provider, shows each account's freshest reading from any of your machines (marked "via" that machine), says the pace in words ("Room to spare", "Runs out early"), and lets you set a budget cap. A login that is signed in only on another machine shows as "Not signed in here"; smart balance and budget caps ignore it. ## Troubleshooting > [accordion] OpenCode won't launch on macOS > > If OpenCode is blocked by macOS Gatekeeper, its diagnostics report a quarantined binary and the exact fix — clear the quarantine attribute with `xattr -d com.apple.quarantine ` (the card shows the full path), then re-check. > [accordion] An agent shows as not signed in > > Run the login command the card suggests, then hit **Refresh**. The sign-in terminal runs ADE's copy of the agent, and the login is shared with your own CLI. > [accordion] An agent shows as downloading or not downloaded > > ADE downloads each agent the first time the machine uses it. The provider's page shows the progress. If the download failed, choose **Retry download**. "Not available on this machine" means the vendor publishes no build for this platform (for example Windows on Arm); installing the CLI yourself does not change that, because ADE only runs its own copy. > [!WARNING] > > Provider calls go directly from your machine to the provider or local endpoint you configure. Keep keys out of committed files — ADE stores them in local encrypted settings. > [card] [Permissions](https://ade-app.dev/docs/configuration/permissions.md) > > Lane scope, provider modes, and how secrets are handled. > [card] [Settings](https://ade-app.dev/docs/configuration/settings.md) > > Every Settings section and a first-pass order. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.