Connect Codex or Claude Code to Studio

Connect Codex or Claude Code to StudioPlugin Studio connects to the official coding engine installed on your computer. You sign in with your own provider account; the engine supplies the model and communicates with its inference service. Studio supplies the plugin workspace, conversation interface, SDK guidance, draft tools and preview. The Plugin website does not run these conversations or provide a shared AI account.The current local integration uses provider-documented connection mechanisms. Its authentication and usage design matches the permitted local-engine patterns described below, subject to the applicable account terms and product-distribution requirements. This is a documented technical assessment, not an OpenAI or Anthropic certification or a guarantee for every deployment. Official guidance was checked on October 13, 2026.Connect your own engine

Install an official Codex desktop app or Claude Code engine on this computer. Studio can discover a compatible engine from Codex, ChatGPT or Claude Desktop, or a standalone CLI on your shell PATH. For Claude Desktop, open its Code tab once so its local engine finishes installing. A chat-only installation is insufficient.

Sign in through the provider's own flow. Studio's Sign in starts Codex's browser authorization or runs claude auth login. You can also authenticate in the official CLI first. Keep your password, API key and account tokens in the provider's own setup; the Plugin website needs none of them.

Open Create with AI, select Codex or Claude above the composer, then open AI tools and skill → Refresh AI tools if the installation or login has changed. Connected means the local engine reports that it is ready. Discovery checks protocol support rather than imposing a fixed version number.

Choose a model and permission mode, describe your plugin and send the message. Model availability comes from the engine's catalog; completing a request still depends on your account access and remaining usage.

Inspect the checked Dock preview, refine it and add the result to Studio. Export a .boltplugin for the consumer PexarBolt app. AI authoring approval does not grant the installed plugin permission to access files, network services or a Dock.

No separate API key is required by Studio when the official engine can use your existing plan. If you configured API-key or cloud-provider access in the engine, that configuration has its own billing. Installing Studio does not include a Codex or Claude subscription.What happens when you send a messageThe local connection has five stages:

Stage

Responsible component

What it does

Request

Studio on your computer

Combines your request, selected attachments, current draft and plugin SDK/skill guidance.

Agent session

Official local engine

Maintains the conversation and requests model inference using its own configured authentication.

Inference

Selected provider or configured inference service

Processes the context and returns model responses under that account's terms and data settings.

Tools

Engine and Studio

The engine can ask to inspect or edit the draft, run checks, use native tools or research a requirement. Studio handles its own draft tools and presents supported approval requests.

Result

Studio

Saves revisions and displays the sample Dock preview. You decide when to export, install and authorize the plugin.

“Local engine” describes where the agent process runs. It does not mean offline model inference: requests, source supplied as context, attachments and tool results can reach the selected provider. Native tools, configured MCP servers and other integrations may contact additional services. Submit material you are authorized to share and use the provider's own privacy, retention and organization settings.Studio saves its draft source and visible transcript locally. The engine may also keep its own session records. Switching providers sends current source and recent visible requirements to the newly selected provider, while keeping their session IDs separate. It does not migrate one provider's credentials to the other.How Codex connectsStudio starts the detected codex app-server as a local child process and exchanges JSON-RPC messages over stdin/stdout. It identifies itself in the protocol handshake, obtains the model catalog and opens or resumes a thread. Plugin draft tools are added to the engine's native tools; responses, tool activity and approval requests stream into Studio.Codex manages its account login and token refresh. Studio uses the account protocol to check readiness and start official browser login; this adapter does not copy Codex credential files or forward its OAuth token to a PexarBolt model proxy. Sandbox and approval choices go through Codex's own controls. Studio's Approve for me delegates eligible requests to native automatic review.OpenAI documents app-server as a product integration interface. Its current authentication guidance allows existing local or open-source applications to continue using that authentication, recommends migration to Sign in with ChatGPT, and excludes commercial or hosted services from app-server authentication. This assessment covers Studio's existing local connection; a commercial or hosted offering must use an appropriately authorized route, such as Sign in with ChatGPT, rather than assuming local authentication permission extends to it.How Claude Code connectsStudio uses Anthropic's official TypeScript Agent SDK to drive the detected, unmodified local Claude Code executable. The adapter selects that executable explicitly, loads its configured settings, preserves the native agent workflow and adds Studio draft tools through an in-process MCP server. It resumes the engine's own session on later turns.Studio's sign-in action invokes the executable's claude auth login. Authentication completes through Anthropic's flow and remains with Claude Code. The adapter does not patch the binary, collect subscription tokens, replace inference endpoints, remove configured authentication methods or silently inject an API key. You can keep your own subscription, API-key or supported cloud-provider configuration in the official engine.Anthropic's product integration rules permit users to authenticate to unmodified Claude Code in another product, with direct user billing and preserved authentication methods. They also require the product operator to agree to Commercial Terms unless otherwise agreed. The current implementation matches those technical conditions; code inspection cannot establish whether a separate operator agreement has been accepted.This is distinct from offering an application's own Claude.ai login or routing users' subscription tokens through another harness, which Anthropic restricts. Its Agent SDK overview directs developers building their own agents toward API-key authentication unless approved. Its current plan notice supports plan usage for the Agent SDK and third-party apps, and notes monthly API credits for eligible plans. That billing notice does not waive the product or credential rules.Accounts, usage and permissionsEach user supplies their own authorized account. Studio does not sell model access, pool subscription accounts or purchase provider usage for users. Requests count under the authentication configured in the engine. Limits, extra usage, credits, model availability and organization restrictions remain controlled by the provider; a Connected status is not a free-usage or billing guarantee.AI tool approval and plugin runtime permission are separate. Codex exposes its native sandbox/reviewer choices. Claude exposes native permission rules and an explicitly selected bypass mode. Studio presents supported requests and checks draft edits; exporting or installing a plugin does not turn those AI approvals into network, filesystem or device grants.The OpenAI terms and Anthropic consumer terms still govern applicable personal use; business/API access and product distribution have their own agreements, including Anthropic Commercial Terms. Use your own account, respect usage limits and content rights, and follow applicable organization and regional access rules. Studio does not change those obligations. PexarBolt is an independent product; use of official interfaces does not imply provider endorsement.Use Codex or Claude Code outside StudioYou can keep the conversation in the official CLI, app or editor and connect it to Studio's Agent service using MCP. In this direction, the provider's client calls Studio's local authoring tools; Studio does not initiate model requests. Agent service is optional for an in-Studio conversation.Enable Agent service in Studio, copy its HTTP or stdio configuration into your trusted MCP client, reconnect and call plugin_context. The service's localhost bearer token grants access to Studio projects; it is not an OpenAI or Anthropic credential. Keep it private. The standalone tooling offers another MCP option without a running Studio app. See external assistant connection for the configuration and tool workflow.If connection is not ready

Symptom

Next step

Not installed

Install an official local engine, or finish downloading the Code engine in the desktop app, then refresh AI tools.

Not supported

Use an installation exposing the required local protocol. Update the provider app/CLI if needed; another discovered installation may already work.

Sign-in needed

Finish the official browser/CLI sign-in, then refresh. Do not paste tokens into a draft or website form.

Connected, but a request fails

Check provider usage, account/model access, organization restrictions and network connectivity. Studio cannot override these.

Subscription versus API billing is unclear

Inspect the official engine's authentication and your provider usage settings before generating. Studio does not switch you to a paid API fallback.

The website publishes setup guidance and downloadable tools. Creating a plugin, publishing a package and signing into an AI provider are three separate actions; connecting an AI engine does not publish your drafts.