AI Assistants and Skills
The Atlas plugin lets a compatible AI assistant use Atlas’s product knowledge and work with your Unity project. It includes tools for inspection, API lookup, source validation, previews and compilation feedback, plus three skills that guide common Opsive workflows.
This connection is separate from choosing Codex or Claude Code as the model inside Atlas. Your external assistant can write code itself and ask Atlas to validate it, or delegate generation to the provider selected in the Unity Atlas window.
Download the connector
Download Atlas AI Connector 0.5.0 for Codex and Claude Code.
The ZIP includes the ready-to-run MCP connector, all three skills, a setup helper, and a README with installation and removal instructions. It is a separate download from the Unity .unitypackage; you do not need the Atlas source repository or a plugin marketplace listing.
Extract the complete ZIP to a permanent folder such as C:/Tools/AtlasConnector. Keep that folder after setup because the assistant runs the connector from there.
Requirements
- Atlas 0.4.0 or newer in Unity, with Connect an AI assistant available. Connector 0.5.0 uses the same local bridge protocol introduced in Atlas 0.4.0.
- A supported product installed in the project and authorized on your Opsive account: Behavior Designer Pro, State Designer, or Ultimate Character Controller Version 3, including its supported controller-family variants.
- Codex or Claude Code running locally on the same computer as Unity. In Claude Desktop, use a local Code session. Cloud sessions cannot reach the local Unity connection.
- Node.js 22 or newer on the machine running the connector, and Unity 2022.3 or newer for the project connection.
Keep the complete plugin folder, including its bundled dist/index.js file. A source checkout needs the connector build described below; copying only the skill files does not provide the connection.
Set up Codex
Open a terminal in the extracted connector folder and run:
node setup.mjs codex
This registers the atlas MCP server for your user account and installs the three skills in ~/.agents/skills. Restart Codex and start a local session for your Unity project. Codex desktop and CLI share the MCP configuration on the same host.
Set up Claude Code
Open a terminal in the extracted connector folder and run:
node setup.mjs claude
This registers the atlas MCP server in user scope and installs the three skills in ~/.claude/skills (or your configured CLAUDE_CONFIG_DIR/skills). Restart Claude Desktop and start a local Code session for your Unity project, or start a new Claude Code CLI session.
If setup cannot find the client command, use --cli followed by the full path to its executable, or use the manual configuration below. Setup preserves an existing atlas server and refuses to overwrite edited skills. The bundled README explains upgrades and removal. Add --dry-run to preview setup without changing anything.
Connect your Unity project
- Sign in to Atlas in the Unity project you want the assistant to use.
- Open Tools > Opsive > Atlas > Connect an AI assistant.
- Turn on Enable connection and choose Copy connection file path.
- Complete the Codex or Claude Code setup above, or the manual configuration below.
- Give the assistant the copied path and ask it to connect using
atlas_connect. Share the path, not the contents of the connection file. - Ask it to check
atlas_get_capabilitiesand confirm the project, installed product version and product access before starting work.
Connections are per project. Select the intended objects in Unity before asking about a graph or character setup. Wait for compilation and reloads to finish before inspecting or applying changes. The connector rereads the connection file after reloads so it can reconnect with the current credential.
Account and authentication
An Opsive account and access to the relevant product are still required. Sign in through Atlas in Unity using the normal browser authorization flow. The connector authenticates locally to Unity, and Unity uses its existing Atlas session for server requests and product-access checks. There is no second Opsive login in Codex or Claude Code for this mode. The assistant does not receive your Opsive account tokens. The connection file contains a separate local credential and must remain private.
Manual MCP configuration
Run node setup.mjs codex --skills-only or node setup.mjs claude --skills-only in the extracted folder. This installs the skills and prints the absolute Node executable and connector paths for your machine.
In Codex, add a STDIO server named atlas under Settings > MCP servers using those values. For Claude Code, merge the following entry into the mcpServers object in ~/.claude.json, preserving existing settings. Close the client before editing its configuration and restart afterward:
{
"mcpServers": {
"atlas": {
"command": "C:/Program Files/nodejs/node.exe",
"args": ["C:/Tools/AtlasConnector/dist/index.js"]
}
}
}
Replace both example paths with the values printed by setup. This is an example entry, not a replacement for the whole configuration file. After the server starts, connect the Unity project using the copied connection-file path. The connector does not require Unity CLI.
Use the skills
| Skill | Useful requests |
|---|---|
build-opsive-behavior |
Create or modify a task, action, condition or ability; check existing functionality first |
diagnose-opsive-problem |
Investigate a selected graph, transition, character setup or unexpected behavior |
review-opsive-code |
Review existing source against the installed product’s APIs and lifecycle |
For example: “Use Atlas to explain why this selected State Designer transition never fires,” or “Use Atlas to review this UCC3 ability’s start and stop behavior.” Skills guide the workflow; they do not grant product access or permission to apply changes.
Inspect, validate and apply
The assistant can inspect selected-object snapshots and look up the installed product’s public and protected API signatures. Documentation searches use the requested product version. Missing exact documentation or contracts remain missing; an older catalog match is identified as approximate and requires explicit opt-in. A Play Mode snapshot is not an execution history.
For a code change, the assistant validates unchanged source and stages a preview with the original and proposed contents. Applying the preview follows the existing Atlas access setting. An intervening file edit makes the preview stale; the assistant must read the new contents and create another preview. The connector cannot grant itself approval.
The apply operation reports compilation status and relevant diagnostics. Generated files are not automatically applied, and a successful compile does not prove the gameplay behavior. Test the result in the intended graph or scene. This connector does not provide arbitrary scene-editing or Play Mode commands.
If generation is delegated to Atlas, it uses the selected Atlas provider unless an explicit model selection is supplied. This can consume that provider’s usage. Knowledge lookup and structural validation do not call a model, although your external assistant’s own model use may still incur charges.
Recover or disconnect
If a response is late or interrupted, ask the assistant to check the existing operation instead of starting another generation. Unchanged retries use the same request ID. An interrupted code-generation request may be recovered; an interrupted advice conversation remains unresolved. Applied-file journals preserve original bytes for recovery, but cancelling an operation does not undo files that were already applied.
Turn off Enable connection to stop bridge access. The credential-rotation control replaces the local credential; assistants that can still read the connection file can reconnect. Keep connection files private. See Privacy and project data.
Knowledge access without Unity
The connector also supports standalone account access for capabilities, knowledge searches, exact published API contracts and structural validation. It cannot inspect a local project, apply files or report Unity compilation in this mode. Supply the product version explicitly; hosted API contracts must have been published for that exact release.
From the plugin folder, run:
node dist/index.js login --api https://atlas.opsive.com --product behavior_designer
node dist/index.js --api https://atlas.opsive.com
Use the first command’s one-time Opsive authorization link. The second command is the stdio server command to configure in your assistant. For another product, use state_designer or ultimate_character_controller. The account creates a separately revocable Atlas AI connector device. Use node dist/index.js logout --api https://atlas.opsive.com to sign it out.
Building the plugin from source
In an Atlas source checkout, run pnpm install --frozen-lockfile, then pnpm agent:package. The complete plugin is assembled under plugins/atlas. Run node scripts/test-agent-plugin.mjs to verify the bundle and connection protocol. The repository’s docs/AIAssistantIntegration.md contains developer setup, API references and release checks. Building a plugin locally does not deploy the matching API service.