# Abacus AI Agent Abacus AI Agent is a coding agent for VS Code that works with any model. Hand it a task and it reads your project, plans, edits files and runs terminal commands, showing every edit as a diff. Run it on one Abacus.AI plan with every frontier model, on the ChatGPT subscription you already have, on your own API keys, or on local models. ![Abacus AI Agent in the VS Code sidebar adding input validation to a TypeScript file, with the edit shown as a diff in the chat](https://static.abacus.ai/cli/readme/extension-agent.png) - **Review before it lands.** Every edit shows as a diff in the chat. When your mode asks first, edits open in VS Code's diff editor to accept or reject: side by side, or hunk by hunk inline with `abacusAgent.inlineDiffEditor`. - **Choose how much it does alone.** Auto (default), Supervise, Accept edits, Plan and Unsupervised. Switch mid-conversation, and set allow, ask and deny rules. - **Plan first when it matters.** Plan mode researches without touching files and waits for your approval before it implements. - **Rewind and fork.** Checkpoints let you take the conversation and your code back to any earlier message, or branch a new conversation from it. - **Keep working while it works.** Queue a follow-up or steer the current turn, and let it run dev servers in background shells. - **Your context.** `@` file mentions, your editor selection, attached images, and your project and personal `AGENTS.md`, loaded automatically. - **Extensible.** MCP servers with OAuth sign-in, and reusable skills you can search and install from settings. - **Fits your layout.** Keep the chat in the secondary sidebar, the activity bar or an editor tab, with searchable history. - **Speaks your language.** The interface follows VS Code's display language, in English and 14 others. ## Use any model, including the subscription you already pay for - **One Abacus.AI plan.** Claude, GPT and Gemini families plus open models such as Kimi, Qwen, DeepSeek and GLM. RouteLLM can pick a model for each turn. - **Your ChatGPT subscription.** If Codex is signed in with ChatGPT, the extension detects it and offers those models. No setup, no credentials copied, and the picker tells you it spends that subscription. - **Your API keys.** OpenAI, Anthropic and Google keys, saved from the settings page to the encrypted credential file the extension and the CLI share. - **Local and OpenAI-compatible models.** Ollama, LM Studio, vLLM, Groq, Together, OpenRouter, DeepSeek or any OpenAI-compatible endpoint. With a local provider, the agent loop runs on your machine and nothing about the turn reaches Abacus.AI. Switch models mid-conversation, even across providers. Each conversation remembers its model. ## Chat launcher The bottom-right **Abacus AI** status-bar launcher opens the configured chat host. Disable it in General settings or with `abacusAgent.showStatusBar`. ## Conversations and tabs When chat lives outside the primary sidebar, that sidebar keeps conversation history. Selecting a conversation focuses its existing tab or opens it in the configured host. **New Conversation** in an editor tab opens another independent tab; closing or moving a tab does not close the other conversations. Use VS Code's editor move/window actions to move tabs; the `window` placement opens a side editor group, not a detached window. ## Review a plan The plan approval card can be resized. **Open plan in editor** opens the exact plan snapshot as editable Markdown with a full VS Code preview; use **Save As** to keep a `.md` copy. **Copy raw Markdown** preserves the source instead of copying rendered text. Use **Source lines** to cite a line in **Feedback for further planning**, then choose **No, keep planning** to send the citation and feedback without approving implementation. The plan stays available in conversation history after rejection and reload, including plans originally read from a file. Historical line references can be copied into a later message. Editing or saving a preview does not approve the plan or change its snapshot. ## Get started 1. Install the extension. It is also on Open VSX for VS Code-compatible editors. 2. Open the **Abacus AI** view and sign in with your email or an Abacus.AI API key. Need an account? See [pricing](https://abacus.ai/pricing). 3. Pick a model, or add your own keys and local models under **Settings → Providers & models**, then describe a task. The **Get Started with Abacus AI** walkthrough on the Welcome page covers the rest. ## Also in your terminal The `abacusai` CLI is the same agent with the same account and conversation history. Run it in VS Code's integrated terminal and it connects to this window: it follows your active file and selection, reads language-server diagnostics and opens edits awaiting your approval in VS Code's diff viewer. It also runs headless (`abacusai -p`) for scripts and CI. See the [CLI install guide](https://abacus.ai/help/abacusai-desktop/cli-installation). ## Privacy Your code is not used for training. Abacus.AI has zero data retention agreements with model providers, and data is encrypted in transit. With a local provider, nothing about the turn reaches Abacus.AI. See the [privacy policy](https://abacus.ai/privacy) and [terms of service](https://abacus.ai/terms). ## Docs and support - [Documentation](https://abacus.ai/help/abacusai-desktop/introduction) - [FAQ](https://desktop.abacus.ai/desktop_faq) - [Pricing and sign-up](https://abacus.ai/pricing) - Support: support@abacus.ai ## Requirements - VS Code 1.105 or later, on Windows, macOS or Linux. Works in remote windows over SSH, WSL, Dev Containers and Codespaces. - An Abacus.AI account to sign in.