Gemini Code Assist

You will get a competent agent inside the editor you already use, and on the Enterprise tier it learns from your private repositories, the trait that matters on a large codebase where public models have never seen your conventions. Day to day it is completions, chat and an agent that edits and runs commands from the same panel.

You will not get git operations or a terminal-only mode, so it stays an editor feature rather than a workflow. Pick it if Google Cloud is already your vendor and the invoice exists. Pick GitHub Copilot if GitHub is, where the invoice is already approved.

The worst line is in the docs: there isn't an option to undo changes made to resources outside your IDE in agent mode. Combine that with yolo mode in VS Code or Auto-approve in IntelliJ and a tool call reaches a database or a cloud resource before you read it.

The consequence is that the agent's blast radius is whatever the developer's credentials reach, and the docs are telling you that plainly. Keep yolo mode off on any machine with production access, and scope the credentials the IDE can see. What it does right: by default every tool use waits for your approval, so the dangerous setting is opt-in.

The documented loop is plan, request a tool, apply the edit, repeat. Two properties stand out. 1. Recitation, the source-matching check available for completions, is not available in agent mode, so the path that writes the most code has the least provenance checking. 2. Context comes from the open project; no repository map or index is described, so retrieval is whatever the model asks to open.

No benchmark is published. The consequence is that verification is delegated to the developer's test suite and the developer's eyes. The observation: the safest mode is the one that edits least.

Formerly Duet AI for Developers, renamed once already, and now narrowed: the free individuals edition stopped serving requests on June 18, 2026, with consumers sent to Antigravity. That is Google choosing where the growth story lives, and it is not here. What remains is a seat business attached to Google Cloud contracts, sticky and boring, sold by a sales force that already has the account.

Pricing power comes from the bundle, not the product, which means price rises arrive as contract renewals rather than announcements. No acquirer, the parent is the acquirer. Position: safe to buy, unlikely to surprise you, unlikely to lead.

The demo edits files in IntelliJ. Procurement: Standard is $19 per user per month on an annual commitment, $1,140 for sixty, or $1,368 at the $22.80 monthly rate; Enterprise is $45 annual, $2,700 for sixty. Billing rides the Google Cloud account, so the invoice needs no new vendor.

No headless mode, so CI gets nothing, and the tool's value is entirely in the editor sessions of the engineers who use it. Onboarding is an extension install. Approved with conditions: annual only, and measure agent use per seat before renewal, because the seats nobody uses are the ones that cost.

Proprietary, Gemini only, no Ollama, no source. The one crack in the box is geminicodeassist.geminiApiKey: drop a Gemini or Vertex key in settings.json and you keep going past the quota on your own bill. MCP servers go in ~/.gemini/settings.json for VS Code and an mcp.json in the IntelliJ config directory, the standard command and args shape.

That is the whole surface I can touch: one key, one server list. I cannot read a line of it, cannot swap the model, and cannot run it without the network. Fine as an extension I use; nothing I would build on.