Large Language ModelsGenerate imagesGenerate videos
Best MCP Servers for Gemini CLI and Antigravity in 2027: What Earns a Slot
Google retired Gemini CLI for individual accounts in June 2026 and moved everyone to Antigravity CLI, but your MCP servers still decide how useful the agent is. See which servers earn a place, with working JSON for both clients, setup traps and safety rules.
On June 18, 2026, Gemini CLI stopped serving requests for Google AI Pro, Ultra and free-tier individual accounts, and the terminal agent most developers open now is Antigravity CLI. Your muscle memory moved. Your tool belt does not have to. An agent is only as useful as the servers plugged into it, and the Model Context Protocol (MCP) is the plug.
Pick the wrong servers and you get a sluggish agent that burns context on tool schemas it never calls. Pick the right ones and it opens pull requests, reads current docs for a library released last month, profiles a slow page and queries your staging database without leaving the terminal. This list sticks to servers with official or widely adopted maintainers, documented setup for Gemini CLI or Antigravity, and one clear daily job.
💡 Short answer: install GitHub, Context7 and one browser server first. Add a database or backend server only when a real task needs it, and keep write access off until you trust the workflow.
What Changed for Gemini CLI Users
The June 2026 Switch
Google announced in May 2026 that it was moving its terminal experience from Gemini CLI to Antigravity CLI. From June 18, Gemini CLI stopped serving individual accounts on the AI Pro, Ultra and free tiers. Enterprise customers on Gemini Code Assist Standard or Enterprise licenses, and Google Cloud projects with paid API access, keep working. The open source repository stays under Apache 2.0.
The new CLI is written in Go and keeps Agent Skills, Hooks, Subagents and Extensions, which are now called Antigravity plugins. The good news for this article: MCP is an open protocol, so your servers carry over. Only the config file location and a few field names change.
💡 People in Google's announcement thread reported that Antigravity CLI burns tokens faster than Gemini CLI did. Every MCP server adds its tool list to the context window, so a short server list is now a cost decision, not just a tidiness one.
Where the Config Files Live
Client
Config file
Remote URL field
Handy commands
Gemini CLI
~/.gemini/settings.json or .gemini/settings.json in a project
httpUrl for streaming HTTP, url for SSE
gemini mcp add, gemini mcp list, /mcp
Antigravity CLI
~/.gemini/config/mcp_config.json or .agents/mcp_config.json in a workspace
serverUrl
/mcp opens the MCP manager
Antigravity IDE and 2.0
Same global file
serverUrl
MCP Store, View raw config
Older Antigravity IDE walkthroughs still list ~/.gemini/antigravity/mcp_config.json. Google's current docs name ~/.gemini/config/mcp_config.json as the global file. If you are unsure, open Manage MCP Servers, then View raw config, and the IDE shows you the file it actually reads.
💡 The classic trap: Gemini CLI wants httpUrl for streaming HTTP servers, while Antigravity wants serverUrl. Paste one config into the other client and the server simply never appears.
Both clients let you trim what a server exposes. Gemini CLI has includeTools and excludeTools. Antigravity has disabledTools and a disabled flag. Use them from day one.
Coding Servers Every Project Needs
Server
Runs as
Best for
GitHub
Remote URL
Issues, pull requests, code search
Context7
Local npx
Current library docs
Chrome DevTools MCP
Local npx
Performance, console, network
Playwright MCP
Local npx
Scripted browser flows
Sentry
Remote URL
Production error triage
MCP Toolbox for Databases
Local npx
SQL and NoSQL access
Supabase
Remote URL
Postgres, auth, storage
Firebase
Local npx
Firestore, Auth, Hosting
Google Workspace
Remote URL
Gmail, Drive, Docs, Calendar
GitHub MCP Server
If you ship code, this is the one. It lets the agent read issues, inspect pull requests, search code across repositories and open new PRs. GitHub hosts a remote server at https://api.githubcopilot.com/mcp/, so there is nothing to install locally. In Antigravity, add this to your MCP config:
In Gemini CLI, one command does the same job: gemini mcp add --transport http github https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer YOUR_GITHUB_PAT". Prefer a local setup? GitHub also publishes a Docker image, ghcr.io/github/github-mcp-server, that reads GITHUB_PERSONAL_ACCESS_TOKEN from the env block.
💡 Create a fine-grained token limited to the repositories you are working in. One community walkthrough also reports that some store installs hit Docker errors, which a plain remote serverUrl entry avoids.
Context7 for Current Docs
Models trained months ago write code against APIs that have since changed. Context7, from Upstash, pulls version-specific documentation and examples straight into the prompt. It exposes two tools: resolve-library-id maps a library name to a Context7 ID, and query-docs fetches documentation for that ID and your question.
The same block works in both clients, because stdio servers use command and args everywhere. Name the library in your prompt and tell the agent to check its docs first. That one habit removes most invented method names.
Browser and Debugging Servers
Chrome DevTools MCP
Built by the Chrome team, this server lets an agent control and inspect a live Chrome browser. It records performance traces and turns them into insights, reads console messages, inspects network requests, takes screenshots and automates actions through Puppeteer. Google's README even lists Antigravity as an example client.
Ask it why the checkout page takes four seconds to paint and it can record a trace instead of guessing.
Playwright MCP
Microsoft's server drives a browser through accessibility snapshots rather than screenshots, so no vision model is needed. The snapshot is structured data, which makes actions more deterministic. It ships more than 60 tools spanning navigation, typing, tabs, network mocking, storage, PDFs and test assertions. The browser opens in headed mode by default; add --headless to the args for unattended runs.
💡 You rarely need both. Start with Chrome DevTools MCP for debugging, and add Playwright MCP only when you want repeatable click-through flows.
Sentry for Production Errors
Sentry runs a hosted MCP service at mcp.sentry.dev with OAuth sign-in, so the agent can pull an issue, read the stack trace and propose a fix next to your code. It handles issue inspection, project management, the Seer skill, and natural-language search tools such as search_errors, search_traces, search_logs and search_issues. Those search tools need an LLM provider configured on the Sentry side.
Sentry's docs give no Gemini-specific steps, but a remote server needs only a URL: use httpUrl in Gemini CLI and serverUrl in Antigravity, with the exact endpoint from Sentry's setup page.
Data and Backend Servers
MCP Toolbox for Databases
Google's open source MCP Toolbox for Databases connects Gemini CLI, Antigravity and other clients to your data. It supports AlloyDB, BigQuery, Cloud SQL, Spanner and Firestore, plus PostgreSQL, MySQL, SQL Server, Oracle, MongoDB, Redis, ClickHouse, Snowflake and more. Prebuilt tools are one click away in the Antigravity MCP Store. For a manual setup, point the config at a prebuilt profile:
When prebuilt tools are too broad, run npx @toolbox-sdk/server --config tools.yaml and define only the queries you want the agent to call. That is the safest way to give a model database access.
Supabase MCP Server
Supabase hosts its server at https://mcp.supabase.com/mcp. Two query parameters matter most: project_ref pins the tools to a single project, and read_only=true removes every mutating tool. You can also restrict the server to feature groups such as database and docs.
Supabase itself tells you to read its security best practices before connecting an LLM to a project. Take that seriously: point the agent at a development project, never at production data.
Firebase MCP Server
Firebase ships a local server that works with your project and your codebase together. Run it with npx firebase-tools@latest mcp. Its tool groups include Core (projects, deployment, security rules validation), Auth, Firestore, Storage, Cloud Functions logs, Crashlytics and App Hosting.
In the Antigravity IDE, skip the JSON: open the MCP Servers menu in the agent pane, pick Firebase and press Install. It writes the config for you.
Workspace and Creative Servers
Google Workspace Servers
Google publishes official remote servers for Workspace, documented in a codelab for Antigravity 2.0, the IDE and the CLI. There is a universal endpoint at workspacemcp.googleapis.com/mcp/v1, plus separate ones for Gmail, Drive, Docs, Sheets, Slides, Calendar, Chat and People. Each uses OAuth 2.0, so you enable the Workspace APIs and the matching MCP services in a Google Cloud project, create an OAuth client, and paste its credentials:
Start with Calendar or Drive, not Gmail. An agent that can read your inbox can also read instructions hidden in a stranger's email.
Linear, Notion and Figma
Google's docs list more than 76 servers in the Antigravity MCP Store, and three are worth a mention for product teams: Linear for tickets, Notion for specs and wikis, and Figma Dev Mode MCP for design context. Install them from the store (Settings, Customizations, Installed MCP Servers, Add MCP in Antigravity 2.0) instead of hand-writing config, since each one needs its own sign-in.
PicassoIA MCP for Images and Video
Most MCP lists stop at code and data. If your agent also drafts blog posts, product pages or launch notes, it needs a way to produce visuals without you leaving the terminal. PicassoIA offers MCP connections for exactly that. A connection exposes four models: PicassoIA Image for text to image, PicassoIA Image Editor Pro for edits, PicassoIA Video for video from text or an image, and Seedance 2.5 Lite for video with audio.
Jobs are asynchronous. The agent starts a generation, receives an ID and polls until the job succeeds or fails. The connector's tools include generate_image, edit_image, two video generators, get_generation, list_generations, list_models and get_account. Up to five predictions can run at once per account, shared across every connection, so tell the agent to queue video jobs instead of firing ten together. Prompts can reach 4,000 characters, enough for a detailed photographic brief.
You create a connection after signing in at picassoia.com/en/mcp/accounts, where the server URL is shown, so the snippet below keeps a placeholder. Check the pricing page for which plans include MCP connections before building a workflow around it.
In Gemini CLI, swap serverUrl for httpUrl, or run gemini mcp add --transport http picassoia YOUR_PICASSOIA_MCP_URL.
Safe Setup Rules That Matter
A server is code you run with your credentials. These rules apply to every pick above:
Least privilege first. Use fine-grained tokens, read_only=true on Supabase, and a development project instead of production.
Trim the tool list. Set includeTools or excludeTools in Gemini CLI, and disabledTools in Antigravity. Fewer tools means a smaller context bill and fewer wrong calls.
Keep secrets out of the file. Gemini CLI expands $VAR and ${VAR} inside env, and it strips variables matching *TOKEN*, *SECRET* and *PASSWORD* from the inherited environment unless you list them there on purpose.
Leave trust off. In Gemini CLI, "trust": true skips tool confirmation dialogs. Reserve it for servers you wrote yourself.
Split global and project servers. Personal tools go in the global file, project tools in .gemini/settings.json or .agents/mcp_config.json.
Treat fetched text as untrusted. Issues, error messages and emails can carry hidden instructions. Do not combine a read source like Gmail with a write tool like GitHub in one unattended run.
Mind the names. Gemini CLI exposes tools as mcp_{serverName}_{toolName}, and Google advises against underscores in server names because they confuse policy parsing.
💡 Debug in one step. Type /mcp in Gemini CLI or Antigravity CLI to see which servers connected, which tools they expose and which ones failed. In Gemini CLI, gemini mcp list shows the same status from the shell. Remember that Gemini CLI's default request timeout is 600,000 ms, so a hung server can look like a slow one.
Gemini 3.5 Flash is a fast multimodal model on PicassoIA that accepts text, images, audio and video in one prompt. It works well as a second pair of eyes for MCP configs, because a missing comma or a wrong field name is the sort of slip it spots in seconds. Here is how to use it:
Open the Gemini 3.5 Flash page and paste your config into the prompt field, with every token replaced by REDACTED.
Fill system instruction with a role such as: "You review MCP client configs for Gemini CLI and Antigravity. Flag wrong field names such as httpUrl versus serverUrl, missing args and secrets stored in plain text."
Set thinking level to low for quick checks, or high when a config refuses to load and you cannot see why. The default is none.
Lower temperature from the default of 1 to about 0.2 so the review stays consistent between runs.
Attach a screenshot of the error in the images field if your client shows one. The model accepts up to 10 images, each up to 7 MB.
Run it, apply the suggested fix, then confirm with /mcp.
💡 The model can also write a first draft of a tools.yaml for MCP Toolbox if you describe your tables and the three queries you actually need.
Make Your Own Images on PicassoIA
Servers give your agent hands. Images give your content a face. Try the approach used for the photos in this article: describe a scene the way a photographer would, with the subject, the light direction, the lens and the mood. Open PicassoIA Image, paste a prompt, refine the result with PicassoIA Image Editor Pro, then animate your favorite frame with PicassoIA Video or Seedance 2.5 Lite.
Run three prompts today, change only the lighting in each, and see which look fits your site. When you are ready to compare more options, browse every model on Picasso IA and start creating.