CodexPanels

Model Bridge

Any frontier model can think. Codex still builds.

CodexPanels separates identity, imported work, intelligence, and execution. Bring supported agent history into ChatGPT, choose an enabled frontier model through Vercel AI Gateway, and keep Codex as the non-negotiable builder plane.

Work layer

ChatGPT projects + native agent import

Instructions, settings, skills, plugins, project folders, MCP setup, and supported recent chats enter through OpenAI’s own import flow.

Intelligence layer

Your preferred frontier model

Choose from the live Vercel AI Gateway catalog. Use platform-managed access or a properly authorized API credential—not a consumer password.

Builder layer · LOCKED

Codex

Codex keeps the repository, tools, permissions, approvals, diffs, worktrees, execution, and receipts. Changing the thinking model does not change the builder contract.

OpenAI-native continuity

Import from another agent.

Use the ChatGPT desktop app’s import flow rather than giving CodexPanels another provider’s password or scraping a private sidebar. The imported work becomes ChatGPT tasks and projects that CodexPanels can organize visually through explicit project and plugin connections.

An optional Workspace Agent bridge can queue a published agent with a stable conversation key. Today that API acknowledges the run but does not return the finished response, so CodexPanels labels it queued instead of pretending it completed.

Read the official import guide
  1. 1
    Open ChatGPT desktop settings

    Choose Import, or Import other agent setup under General when that is where the rollout appears.

  2. 2
    Select supported sources and items

    Choose setup, projects, and recent work. Existing ChatGPT setup is not deleted.

  3. 3
    Finish connection authorization

    Imported plugins and MCP servers may require sign-in again. CodexPanels shows those as unresolved receipts, not fake green lights.

Vercel AI Gateway

One catalog. Multiple model companies. Explicit cost.

CodexPanels asks the Gateway for the current catalog instead of freezing model names into templates. Workspace owners choose the allowlist, default, routing, failover, budgets, and whether credentials are platform-managed or BYOK.

OpenAIAnthropicGooglexAIDeepSeekMore via Gateway

Connection contract

Consumer subscription ≠ API access.
A Claude, Gemini, or ChatGPT consumer login is not treated as a reusable model credential unless that provider publishes an authorized integration for it.

No provider passwords.
Use Vercel-managed access, scoped provider API credentials, enterprise federation, or a supported OAuth flow. Secrets remain server-side or in the user’s local Codex environment.

Live discovery.
Available models, endpoints, and routing are read from Gateway at runtime; marketplace panels declare capability needs instead of hard-coding a vendor.

The non-negotiable builder

Codex is why the name still works.

The product is not a generic multi-model chat shell. A frontier model may analyze, plan, critique, or draft the Turn Capsule. Codex receives the bounded build instruction and remains accountable for actions in repositories and connected tools.

  • Repository and worktree context
  • Sandbox and approval policy
  • MCP tools, skills, plugins, and hooks
  • Diffs, checks, receipts, and resumable runs
Codex Gateway profile
[model_providers.vercel]
name = "Vercel AI Gateway"
base_url = "https://ai-gateway.vercel.sh/v1"
env_key = "AI_GATEWAY_API_KEY"
wire_api = "responses"

[profiles.vercel]
model_provider = "vercel"
model = "<choose-from-live-gateway-catalog>"

The key is provided through the environment. It never belongs in a panel manifest or pasted chat.

Brand decision

Keep CodexPanels.

Glyphd OS and Polyglyph are broader creation environments. CodexPanels is narrower and more legible: it is the visual operating surface for people whose actual builder is Codex—even when another frontier model supplies a layer of intelligence.

Start onboardingReview connections