Pixis

Pixis is an AI coding agent that works inside Visual Studio Code. It can explore a repository, explain code, propose and apply changes, run commands with your permission, use specialised skills and connect to local or remote MCP tools.

Everything Pixis needs to run is bundled with the extension. There is no separate server to start, no background service to configure and no additional command-line tooling to install.

Requirements

• Visual Studio Code 1.125.0 or newer.
• A Pixis access token — it looks like pixis_ followed by 32 characters.
• A supported desktop platform:
  • macOS Apple Silicon (darwin-arm64);
  • Windows 64-bit Intel/AMD (win32-x64).

Cursor can install Pixis when its VS Code compatibility level meets the requirement.

Linux, macOS Intel, Windows ARM64, Windows 32-bit and browser-hosted VS Code are not included in this release.

Setup

1. Install Pixis from the Marketplace.
2. Open a project folder in VS Code — Pixis works against a workspace, not a single loose file.
3. Select the Pixis icon in the Activity Bar.
4. Open Pixis Settings.
5. Paste your access token into the Pixis access token field at the top of the page and select Sign in.
6. The status line should read Signed in. Pixis is ready to use.
7. Return to the Pixis chat and send a request.

One token covers everything Pixis does — there are no per-service keys to collect and nothing else to sign up for. It is stored using VS Code SecretStorage, is never shown again, and is never included in copied diagnostics. You can sign out from the same panel at any time.

Using Pixis

Ask Pixis for an outcome in ordinary language:

Explain how authentication is wired in this repository. Do not modify files.

Fix the registration validation bug and run the relevant tests.

Press Enter to send and Shift+Enter for a new line.

Responses stream into the conversation as they are generated. Agent operations appear in a compact Working section where commands, file reads, edits, results and failures can be inspected individually.

Build and Plan modes

Use the mode selector in the composer:

• Build investigates, edits files, runs commands and verifies changes.
• Plan investigates and produces a plan without using the edit or write tools.

Supported edit and write operations may open a native proposed diff before permission is requested. Shell and MCP operations show their command or input instead; a proposed diff is not guaranteed for those tools.

Reasoning effort

Use the effort selector beside the mode selector in the composer to choose how much the model reasons before responding:

• low
• medium (default)
• high
• xhigh
• max

The selection applies to the next message and can be changed again for any later turn.

Permissions

Side-effecting operations are controlled by permission prompts:

• Allow permits this occurrence.
• Always stores a matching permission rule for future operations.
• Deny rejects the operation and stops the affected execution path.

Permission-requiring calls are presented sequentially. Denied operations are not retried through another tool.

Use Always carefully because saved permission rules persist across sessions.

Agents

Type @ at the start of the composer or open Pixis Settings → Agents.

• Explore searches and explains the codebase without modifying files.
• General carries out work that is already well enough defined to act on, in one focused turn, then returns control to the normal conversation.
• Architect designs an implementation approach when a change is big or unfamiliar. It reads the code and returns the order of work, the files involved and the trade-offs it weighed, in the conversation. It cannot edit or write anything.
• Plan does the same investigation and writes the finished plan to a markdown file for you to review and approve. It can change nothing except that plan file.

Only General can change your code; the other three cannot.

The selected agent appears as a composer chip and applies to the next message.

Inline and background work

Delegated work runs one of two ways:

• Inline is the default. The task runs and Pixis waits for the result before carrying on — the right choice when the next step depends on the answer.
• Background detaches the task so Pixis can get on with something else, and notifies you when it lands.

Several tasks started together run in parallel either way, so a wide fan-out does not need to be backgrounded to overlap.

Custom agents

Ask Pixis to create a custom subagent (for example, "create a subagent that reviews pull requests for security issues") and it writes a new agent definition using the create-subagent skill, generated for that request rather than a fixed template. Restart the engine from Pixis Settings afterward; the new agent then appears in this same @ menu and in Settings → Agents alongside the built-in ones.

Skills

Type /, select Skills, then choose a skill. Skills can also be selected from Pixis Settings → Skills.

Bundled skills:

• commit — generate a Conventional Commits message from the current diff.
• pixis-proof — explain what a Pixis proof card verifies.
• project-pdf — generate a PDF overview of the current project.
• site-redesign-ui-ux — design, audit or implement production-ready web interfaces.
• create-subagent — write a new custom subagent from a plain-language description, ready to use after a Restart Engine.

Add the task beside the selected skill chip before sending. Skills you add at project or global scope also appear in the catalog.

Slash commands

Type / at the beginning of the composer:

• /skills opens the skill list.
• /mcps opens MCP settings.
• /init runs guided AGENTS.md setup.
• Safe configured commands may also appear.

Commands that contain embedded shell execution or unsupported background-task behaviour are disabled with an explanation rather than run without the expected safeguards.

File and image attachments

Use the paperclip button to attach supported files.

Images:

• PNG
• JPEG
• GIF
• WebP

Text and code attachments include common source, configuration, markup and data formats such as TypeScript, JavaScript, Python, Go, Rust, Java, C/C++, Markdown, JSON, YAML, TOML, XML, HTML, CSS, SQL and shell scripts.

Each selected file may be up to 20 MB. Text sent to the model is bounded to 50 KB, 2,000 lines and 2,000 characters per line. Pixis clearly marks truncated content and rejects unsupported binary formats.

MCP servers

Open Pixis Settings → MCP or use /mcps.

Pixis supports:

• local command-based MCP servers;
• remote URL-based MCP servers;
• project or global configuration scope;
• local environment variables;
• remote request headers;
• configurable timeouts;
• remote OAuth sign-in and sign-out;
• connect, disconnect and permanent removal.

After a server reports connected, ask Pixis to use it:

Use the ms-learn MCP server to search Microsoft documentation for TypeScript configuration.

MCP tools, server instructions, resources and resource templates are integrated into the agent's available context.

Do not commit secrets in project MCP configuration. Prefer environment variables or OAuth where available.

Sessions and history

Pixis stores conversation history locally. Use the session drawer to:

• return to earlier conversations;
• see running sessions;
• start a new session;
• switch between sessions;
• remove a session where supported.

Todo lists remain attached to the assistant message that created them. Completed tool groups collapse automatically, while failures stay visible.

Copying content

Copy buttons are available for:

• complete assistant responses;
• user messages;
• command and tool details;
• expanded tool output and errors.

Settings

Pixis Settings opens in a wide editor view with pages for:

• Pixis access token — sign in, check status, sign out.
• MCP
• Agents
• Skills

Behaviour switches live in ordinary VS Code settings:

These take effect on the next engine start — run Pixis: Restart Engine after changing one.

Troubleshooting

Pixis asks you to sign in, or your token is refused

Open Pixis Settings and check the token begins with pixis_ and runs to 38 characters. A refused token may have been revoked or copied incompletely. Sign in again, and restart the engine if requested.

The engine does not become ready

Run Pixis: Restart Engine from the Command Palette. If the problem continues, run Pixis: Copy Engine Diagnostics and include the result in a support request.

Skills or agents are missing

Reload the VS Code window, restart the Pixis engine and reopen the project folder. Confirm the installed Pixis version is current.

Windows reports that Pixis is unavailable

Confirm VS Code is version 1.125.0 or newer and reports x64. This release does not include Windows ARM64.

A newly published update is not installed

Reload VS Code and run Extensions: Check for Extension Updates.

Privacy and security

• Your Pixis access token is stored in VS Code SecretStorage.
• The bundled engine listens only on localhost and uses generated authentication.
• The webview cannot access the filesystem or engine directly.
• Prompts, selected file content and tool results are sent to the model service Pixis uses on your behalf.
• MCP tool input is sent to the MCP server selected for that operation.
• Pixis does not display an estimated price meter when the engine cannot provide reliable billing data.

Review your organisation's source-code and MCP data policies before using Pixis with confidential material.

Support

Use the support channel supplied by the Pixis project owner or your organisation's deployment administrator. Include the Pixis version, editor version, operating system, exact error and Pixis: Copy Engine Diagnostics output. Never include credentials or private source code.

Release notes

Release history is available in the Marketplace Version History section and in the CHANGELOG.md included with the extension.