Your agent: running LiveOps end-to-end

Last week the MCP server was read-only. Now an agent can edit and manage your configurations, run experiments end to end, and build dashboards to watch them. It asks first, every time.

Preview before save. It shows every value it would change, before and after, and a change is refused if a teammate edited the configuration since the agent read it.
Experiments, start to finish. It sets up groups and splits, reads results with lift and confidence, and ends an experiment only when it names it exactly.
Writes are opt-in. A connection stays read-only unless you allow changes when signing in. Existing connections don’t change.

Heads up: approving an agent change to your primary configuration pushes it to live servers right away, just like the dashboard.

Read the MCP guide

Ramp experiments up gradually

An experiment no longer has to reach every eligible player on day one. Turn on Gradual rollout when you create it, pick a starting share like 5%, and add steps such as “25% after 1 day” and “100% after 3 days”, or choose a preset. Every group grows together, so the comparison stays fair all the way up.

Once it’s running, the experiment page shows the ramp and the next step. Advance it early, pause it, set a share yourself, or reshape what’s ahead. Lowering or pausing only stops new players from joining: anyone already in keeps their group. Your agent can do the same, with your approval.

Set up a rollout

Three weeks until legacy experiments end

A reminder, not news: legacy experiments shut off on October 21, so move anything you rely on to the current experiments surface. On Roblox, that needs SDK 1.0.0, which releases early next week. We’re taking extra time to make sure everything runs smoothly with the new capabilities this build brings. The Experiments guide is rewritten to help, with step-by-step migration and a live demo of how players get assigned.

Read the docs

Your agent now knows how to operate your game

The MCP server now ships the playbooks our own assistant uses and can read these docs, so your agent does things the way we would instead of guessing.

A skill for each area. Queries, cohorts, dashboards, configurations, experiments, deeplinks, and Community Pulse. Agents load them with load_skill; they hold instructions, never project data.

It reads the docs. search_docs and read_doc cover the guides, SDK and API reference, and changelog, so it looks things up instead of answering from memory.

More query tools. If your role can view queries, agents can now build insights, funnels, and retention on the fly, evaluate cohorts, and query sentiment.

See the skills

Smaller things worth knowing

Full API reference. 92 documented endpoints, up from 36, with a playground that calls the API from your browser. Browse it.

MCP sign-in fixed. Connecting from Cursor and VS Code works again.

Custom roles hold for agents. Chat and MCP enforce exactly what the API does.

Also. Revoked SDK keys stop working within moments, and Roblox markers with an empty table are accepted.

On the horizon

early access a web SDK

When we added Unity in issue 01, we said web was on the list. It’s here in early access: a web SDK, so browser games and web builds get the same configs, experiments, and analytics as Roblox and Unity, in the same dashboard. It runs in the browser and on your servers, and the full reference is already live. Want in? Reply to this email.

Your agent: running LiveOps end-to-end