Skip to content
| Marketplace
Sign in
Visual Studio Code>Machine Learning>Cenmax AINew to Visual Studio Code? Get it now.
Cenmax AI

Cenmax AI

Cenmax

|
4 installs
| (0) | Free
OAuth-first Cenmax AI coding agent for Visual Studio Code
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

Cenmax AI for Visual Studio Code

Cenmax AI brings the OAuth-first Cenmax AI coding agent into a workspace-scoped VS Code chat view. It streams answers, can inspect and edit workspace files, runs bounded shell commands, keeps long-running background terminals, fetches public references, and supports safe mid-run steering.

Run and steer

While a turn is running, the composer shows a red Stop button that interrupts the current model request or tool call (also available as Cenmax AI: Stop Current Turn). The run's reasoning is shown as a collapsible Thinking · ~N tokens card, where the token figure is a live estimate while the response is generated.

Background terminals

The agent can keep servers and watchers alive between turns with the terminal tool, for example php artisan serve, npm run dev, or a queue worker. It starts a command, lists open terminals, reads incremental output, and stops a terminal when it is no longer needed. Background terminals belong to the workspace and are terminated when the extension host unloads; use Cenmax AI: Stop Background Terminals to stop them all from the Command Palette.

Open files and attach images

The agent can show you a file with the open tool: it opens a persistent editor tab, optionally at a line, or reveals the item in the Explorer file view. Opening is read-only and available in every mode.

Paste a screenshot or image directly into the composer to attach it to your next message (PNG, JPEG, WebP, or GIF; up to four images). Images are downscaled and re-encoded before they are sent, previewed above the composer, and shown in the conversation. Use Enter to send; the attached image can be removed with its × control before sending.

Account tools and skills

When you sign in with OAuth, the extension loads the MCP connections and Agent Skills that are enabled on that Cenmax AI account:

  • Connected MCP tools appear to the agent automatically. Each call runs server-side through your account workspace with an idempotency key, so the extension never receives the third-party credential. MCP actions change external services, so they ask for approval outside plan mode; plan mode never runs them.
  • Enabled skills contribute their name and description to the agent's context and are loaded in full only when the agent uses the skill tool. Skill content is treated as guidance, never as permission to bypass workspace safety rules.

This requires the account-level workspace permissions. If your account was connected before these permissions existed, choose Add or reconnect account from the account menu and approve the updated permissions. The API-key fallback cannot use account tools or skills; sign in with OAuth for those.

Run locally

  1. Run npm run build in this directory.
  2. Open this directory in VS Code.
  3. Press F5 and choose Extension Development Host.
  4. Open a folder in the development host, then select the Cenmax AI activity-bar icon.

Use Cenmax AI: Sign In with OAuth from the Command Palette. An API key can be stored as a fallback with Cenmax AI: Set API Key Fallback. Both token types are kept in VS Code SecretStorage.

The account picker in the composer and Cenmax AI: Switch Personal or Work Account let you move between connected personal and company contexts. Each context keeps its own OAuth token in SecretStorage and therefore retains its own billing, subscription, quota, and usage boundary. Use Add or reconnect account to authorize another context in the browser. Account changes are blocked while an agent run is active.

The default model is worker, the default thinking level is low, and the default mode is act. OAuth connects to https://console.ai.cenmax.in, and API requests use https://api.ai.cenmax.in/v1. Configure model, thinking, mode, web search, URLs, and step limits under Settings → Cenmax AI. The two URL settings are application-scoped and restricted to the official Cenmax AI hosts, so a repository cannot redirect account tokens elsewhere.

Safety modes

  • plan: permits inspection and refuses workspace mutations.
  • act: asks before any state-changing tool.
  • auto: runs ordinary project work unattended — builds, tests, formatters, package scripts, workspace edits, local Git work, and known dev servers bound to localhost. An independent command reviewer checks destructive, opaque, or unusual commands against your request; if it cannot confirm them, you are asked. External web fetches, credential-holding files, project-local executables, environment assignments, dynamic cd targets, and dependency installs require review. Command review uses your signed-in OAuth account, so the API-key fallback falls back to asking. Credentials, privilege escalation, publishing, system changes, and network clients are never run unattended.

Messages sent during an active turn are queued and admitted after the current model request or tool call reaches a safe boundary.

Security model

  • Workspace Trust: the extension is not supported in restricted (untrusted) workspaces and does not activate there. In a trusted workspace, Auto mode still never runs privileged, publishing, credential-bearing, or unguarded network commands without review.
  • Approvals show the full command (up to 4,000 characters) rather than a truncated preview, so a destructive tail cannot hide behind a benign prefix.
  • Network access: web fetches resolve the hostname once, reject private, loopback, link-local, CGNAT, multicast, and IPv4-mapped addresses, and pin the validated address for the request so DNS rebinding cannot redirect it. Web fetches require approval in Auto mode.
  • Reads: read asks before opening files that commonly hold credentials (.env, .ssh, key files, and similar); grep skips those files and caps pattern length, file size, line length, and search time.
  • Commands: Auto rejects leading environment assignments, dynamic cd targets, writable-location executables, and copy/mkdir/touch paths that resolve outside the workspace or through a symlink. Terminal servers are accepted only as exact argv forms bound to loopback.
  • Account data: tokens live in VS Code SecretStorage; account tools and skills require OAuth with the workspace scopes, and connected-tool actions are approval-gated.
  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft