FAQ

FAQ

What is Codexify?

A local MCP service that gives a ChatGPT conversation programming tools on your machine. ChatGPT reasons; Codexify runs files, commands, Git, diffs, skills, and configured upstream tools. The private owner interface adds persistent chats and optional browser-assisted goals. See Home.

Does it consume Codex quota or remove ChatGPT limits?

The ordinary ChatGPT Chat workflow does not start a Codex/Work agent. It still uses the selected ChatGPT model's allowance and is subject to OpenAI's account/service limits. Calling another metered service through a tool can incur that service's normal usage. There is no guaranteed unlimited model access.

Is a particular paid plan sufficient?

Custom MCP plugin access and native tunnel permissions must both be available to the account/workspace. They are separate permissions, so an old fixed plan table is not a reliable setup prerequisite. Follow Connecting to ChatGPT.

Do I need Rust or Codex installed?

No Rust toolchain is needed for release binaries. Codex is optional: Codexify can read its existing user config for MCP/project/skill discovery and optionally use the CLI, but explicit configuration works without it. Source builds need Rust; see Installation.

Does it expose my computer publicly?

The recommended tunnel connection is outbound-only, with a local authenticated MCP listener. The owner interface is also loopback-only. External HTTP mode needs your own authenticated ingress. None of these replace OS isolation from a connector user who can run unrestricted commands. See Security model.

Can I switch projects in one conversation?

Yes, using the setup card's Switch to another project. The agent must read the new brief. Running commands stay in their original directory; pending messages retain their destinations. The legacy statement that every conversation must remain bound forever is obsolete. See Workspace selection.

Can I create a worktree based on another worktree?

Yes. Choose Worktrees → New from here, or a validated registered base with the worktree choice enabled. The new checkout starts at that base's current commit, without copying its uncommitted files. Selecting the checkout name itself reuses it. See Worktree isolation.

How do I continue in a fresh ChatGPT conversation?

For manual takeover, use the generated full continuation prompt and continue_task. Verified browser automation handles stale-schema and conversation-length replacement automatically before its next normal submission. It transfers the existing task and retires its old physical owner. The legacy resumePath option only reuses the workspace; it does not transfer chat/process/diff ownership. See Context and memory.

Where are the chat controls, and why does Return queue?

Run codexify chat for the private owner interface. Queue is the default mode; Steer releases input for the current turn, with browser wake-up still subject to its queue/cooldown when offline. Hold the send button to choose either mode for just one message. While the agent waits for your answer, Return and the modified Return send the answer directly. See Agent chat.

Why can a goal be active while the agent is offline?

Goal status is the saved objective state; presence is based on observed tool activity and explicit waits/leases. ChatGPT can be thinking or stopped while the goal remains active. agentChat.offlineTimeoutMinutes controls display and automatic offline eligibility together (default 5; restart to change). The queue/sending/first-tool-call indicators and browser waiting bar show dispatch progress. See Agent chat.

Why does waiting for cooldown not show Retry?

The action has not failed or spent a retry merely by encountering an existing cooldown. The waiting bar shows a gray spinning hourglass, countdown and queue position, with Do now and Cancel even without a goal. Do now overrides Codexify's local cooldown for one attempt; it cannot remove ChatGPT's own limit.

Where is the explanation for a blocked or completed goal?

goal_finish requires a concrete reason plus unblock steps for a blocker, or a work/results/verification summary plus suggested next steps for completion. These reports appear inline in persistent chat history. A new message sent after blocking can wake an offline conversation; earlier queued requests remain private.

Are unread dots shared across browsers?

Yes. Owner read positions persist on the server. Your own messages, tool calls, and receipt changes do not create unread dots. This is distinct from the agent's read cursor.

Why is file materialization unavailable?

It defaults off because ChatGPT may request confirmation for generated downloadable files. Ordinary chat attachments, inline text/images, and local inspection still work. Enable artifactEgress.enabled only when you want native exports and file-resource results, then apply the required restart/schema refresh. See Tools reference.

Why did an update leave older browser behavior?

The server and companion are separate installations/processes. An older protocol-compatible companion can remain running. Update the server, run codexify browser install, then Stop/Start browser in global settings. See Installation.

Why do retries reload an already-open ChatGPT conversation?

A known access/sign-in/no-composer failure needs a fresh load; healthy pages are reused. Rechecking the same failed document would fail forever. Auth/account checks still precede Send. See Browser automation.

Can I pause or dismiss retries?

Yes. Pause retries retains the error; Dismiss clears it and cancels retries. Global Automatic retries controls the policy immediately. The active goal and unsent user messages are not deleted. See Browser automation.

What do Thinking hours and achievements measure?

Thinking hours is the stats card label for estimated active time, not private reasoning or billed usage. Statistics supports Today/1 day through Lifetime; seven days is default. See Analytics.

How many tools are available?

The v2.3.24 default registry has 47 model-visible native tools plus 15 app-only helpers before upstream imports. Experimental agent tickets add ticket_wait; materialization adds an export tool; disabled features, authorization, widgets, and MCP exposure change the surface. Use the live schema and Tools reference, not the old single-project/multi-project counts.

Why are there logs without an explicit audit filename?

New defaults enable audit, redacted command previews, and Info-level payload logging. The audit destination is chosen automatically. Existing explicit opt-outs survive upgrades. Audit logging shows how to disable each surface.

Is Codexify an official OpenAI product?

No. It is an independent MIT-licensed project using MCP and the official tunnel runtime. The browser companion is an unofficial UI integration. Report issues with sanitized diagnostics.

---

Reviewed against v2.3.24 on 29 October 2026. Source: src/config.rs · src/registry.rs.