Connections.

The big one. Epic now ships a first-party Claude Code plugin, and UE 5.8 has an MCP server built into the editor. Hundreds of tools across 30+ toolsets via Unreal's ToolsetRegistry — spawn and arrange actors, edit Blueprints, author materials, build UI widgets, script Sequencer, configure Niagara, wire Control Rig and Gameplay Ability System logic, run automation tests.

Two gotchas, both real. The Unreal editor must be open — no editor running means Claude has no hands. And you may have to enable the Toolset registries manually inside Unreal; skip that and the server connects but exposes nothing useful. Verify with /mcp, then prove it: "list the actors in the currently open level."

The biggest blind spot Claude currently has: it can read our files but knows nothing about the depot they came from. Perforce publishes p4mcp-server, covering files, changelists, shelves, workspaces, jobs, reviews, and streams — which turns "is anyone holding a lock on this asset?" and "what changed in this stream since Tuesday?" into questions with answers.

Read the label. Perforce ships this as a community supported project — explicitly not officially supported by Perforce. It defaults to read-only, does ownership checks, and asks for confirmation before destructive operations, which is exactly the posture we want. Because it needs P4Python and depot credentials, this is an IT-provisioned install, not a self-serve one.

Ask engineering to stand it up in read-only mode first: github.com/perforce/p4mcp-server

Sentry has a first-class Unreal Engine SDK, so crashes from real builds on real machines land in one dashboard. Connecting it promotes log-and-crash-triage from "read the log on my desk" to "read every crash the whole player base hit." Errors, issues, traces, performance data, and triage actions — all queryable in plain English.

Sentry also runs its own marketplace, whose build adds a dedicated subagent Claude delegates to automatically for anything Sentry-shaped — worth it if you live in crash triage:

OAuth — a browser window opens the first time, then it remembers. Try: "what's the top new crash since Thursday's build, and which changelist looks responsible?"

Product analytics with events, session replay, feature flags, and experiments — queryable in natural language or HogQL. This is the other half of balance-data: instead of tuning a data table on instinct, ask where players actually fall off, then tune the table and watch the funnel move.

Sign-in happens on first use and auto-routes to the right data region. Only relevant if we're actually sending telemetry to PostHog — check with production before installing, and if we're on a different analytics stack, ask whether it has an MCP rather than adding a second pipeline.

Atlassian's remote server went generally available in February 2026 with Claude as its first official partner. Search and summarize across Jira issues, Confluence pages, and Compass components, and create or update them — respecting whatever permissions you already have, so it can't see what you can't.

Or connect it directly:

OAuth 2.1 in the browser, or an API token for headless setups. This is what makes spec-to-plan produce tickets instead of a list you then retype.

For whatever lives in git rather than the depot — tools, build scripts, this site. Issues, pull requests, reviews, and Actions runs become things Claude can read and act on, which is how "summarize what landed this week" stops being a manual sweep.

On GitLab instead? Same shape: claude plugin install gitlab@claude-plugins-official. For automatic PR review by @claude mention, see Studio Setup.

Reading a channel's history is the fastest way to reconstruct why a decision was made. Useful for meeting-to-action and risk-radar — "what did the combat channel conclude about stagger last week, and did anyone own the follow-up?"

Work workspace only. Same rule as every connector: sign in with your Lasting Light account, never a personal one.

If your tasks live in Linear rather than Jira, this is the equivalent — read and update issues directly, so status conversations start from the tracker instead of from memory.

No OAuth on your account? A Personal API key works too (Linear → Settings → Security & access).

If our GDDs and process docs live in Notion, Claude reads and updates them directly — which is the difference between a design doc that drifts and one that stays current because updating it costs a sentence.

First time only: open each page you want visible → ••• → Add connections → select Claude. Pages you don't share stay invisible.

HUD and menu mockups, UI flow wireframes. Claude reads a file's real structure — layers, constraints, tokens — rather than guessing from a screenshot of it.

Needs Dev Mode enabled on the file you're pointing at.

Progression flowcharts, quest state diagrams, retro boards — read and edited rather than described back to you.

Then /mcp → sign in. Enterprise-plan boards may need an admin to enable MCP access first.

For blockouts and DCC work before anything reaches the engine. The dominant community project in this space — actively maintained, very widely used.

Two halves. In Blender: install the addon from github.com/ahujasid/blender-mcp (Preferences → Add-ons → Install), enable it, start its connection server from the sidebar. Then on the Claude side:

No API key needed for core use.

Text-to-speech for barks, item descriptions, and dialogue read-throughs before a VO budget exists. Hear the line in context before committing to it.

Key at elevenlabs.io/app/settings/api-keys (free tier: 10k credits/month). Personal key — always add this one with -s user.

Relational content databases — item tables, quest lists, dialogue trees — with real structure instead of a flat sheet that nobody trusts.

Then /mcp → sign in. A Personal Access Token from Airtable → Account → Developer hub also works.

If we're shipping localized, this is the connector that stops loc from being a spreadsheet emergency at the end of every milestone. Crowdin's hosted MCP — rebuilt as v2 in March 2026 — handles branches, file organization, labeling, task assignment, and progress checks conversationally: "which locales are blocked on the new barks, and who owns them?"

Hosted, so nothing runs locally, but it authenticates with a Personal Access Token tied to a real Crowdin account — which makes it a provisioning decision, not a self-serve install. Setup lives at support.crowdin.com/crowdin-mcp-server. Ask production whether we have a Crowdin org before wiring anything.

Real-time web research with sourcing — competitive teardowns, "how do the top roguelites handle death and recovery," in one prompt instead of an afternoon of tabs.

Key at perplexity.ai/account/api/group. Personal key — use -s user.

A second, independent index — useful when you want results that don't all route through one company's ranking.

Free key (2,000 queries/month) at brave.com/search/api. Personal key — use -s user.