Website · Dashboard · API docs
Run A/B tests and personalizations on your own website, from your coding agent.
Mida is an A/B testing and personalization platform. A small JavaScript tag on your site puts visitors into variants and reports conversions. This plugin connects your agent to your Mida account, so you can build a test, launch it, read the result, conclude it, and roll the winner out without switching tabs.
"What's running on the pricing page right now?"
"Create a test for a shorter signup form and preview variant B"
"Has the homepage hero test reached significance yet?"
"Conclude experiment #208, variant B won, we're shipping it"
Gemini CLI
gemini extensions install https://github.com/mida-so/mida-plugin
Codex
codex plugin marketplace add mida-so/mida-plugin
codex plugin install mida-so
Claude Code
claude plugin marketplace add mida-so/mida-plugin
claude plugin install mida-so
Any MCP client, point it at the hosted server:
{ "mcpServers": { "mida": { "type": "http", "url": "https://mcp.mida.so/mcp" } } }Per-client setup instructions live at mida.so/mcp.
The marketplace add commands work because this repo is its own single-plugin
marketplace: it carries .claude-plugin/marketplace.json for Claude and
.agents/plugins/marketplace.json for Codex alongside the per-client manifests.
Listings for Grok Build and Cursor are in review. Once they land, it's
/plugin install mida-so in either one.
On first use your client opens a browser to sign in to Mida with OAuth. There is no API key to paste and nothing is written to your machine.
You need a Mida account and the Mida tag installed on the site you want to test. Ask your agent for the install snippet once you're connected.
MCP server. The hosted Mida MCP server, covering experiments (create, launch, update, conclude, reopen), results and statistics, goals and custom events, hypotheses, exclusion groups, personalization campaigns, project configuration, and team access.
Skills
| Skill | Covers |
|---|---|
mida |
Entry point: picking the right project, reading Bayesian vs. frequentist results, and the order of operations for concluding a test and serving a winner |
mida-experiments |
The workflow: hypothesis, variants, goal, preview, launch, then calling the result |
The skills are here because a few Mida behaviours are easy to get wrong from the tool schemas alone. An experiment's display number is not its API id. A concluded test stops serving even while its status still reads active. Setting a test's status clears a conclusion written before it. Each one is documented with the correct sequence.
| Endpoints called | https://mcp.mida.so/mcp, the hosted Mida MCP server, and the only host this plugin contacts. It reaches Mida's API server-side. |
| Credentials | OAuth to your Mida account, handled by the MCP client. No API keys, no environment variables, no local credential files. |
| Scope | Your own Mida organizations and projects. The plugin reads and writes experiment configuration and reads result data for those projects. |
| Local execution | None. No hooks, no commands, no scripts. One remote MCP server plus Markdown skills. |
| Telemetry | None beyond the MCP calls themselves. |
Mida is A/B testing and personalization in one engine. Testing answers which version works. Personalization answers who it works for. Both run as real variants in a real test, so personalization is something you can actually prove rather than take on faith.
- A/B testing on any site, with a tag that loads in about 20ms
- Mida dashboard for results, goals, and rollouts
- Public API docs
MIT, see LICENSE.