I'm trying to use Codex via the Xcode interface, and I can not find a way to set the reasoning level. There are no settings to be found anywhere. What am I missing? I mean, I can continue to use Codex's own app, but I was hoping to use Xcode's new agentic interface, which seems superior to v26 and somewhat more useful. But if it's going to default to medium and I have no control over it, of course, I would continue using Codex's app rather than Xcode. What's weird, I get responses like this when I ask about: What's your current reasoning level and how can I adjust it? This session appears to be using the default reasoning level: medium. You generally adjust it outside my reply, in the model/session settings for Codex or the API request. Typical options are: • low: faster, cheaper, best for simple edits or direct questions • medium: default balance • high: better for complex debugging, architecture, tricky code changes • xhigh: deepest reasoning, slower, useful for hard multi-step work Inside this chat, I can’t change my own base reasoning level mid-response. You can still tell me “use high reasoning for this” and I’ll treat the task more deliberately, but the actual runtime setting has to be changed in the client/session configuration. So, it doesn't even know the level it's using or how to adjust.

In the video "Xcode, Agents and You" (direct link), the presenter asks for a walkthrough of the open project and Xcode presents a rich-text overview of the app's architecture, with color-coded diagrams and clickable links to the files it discusses. I can't get that to work when I try it, though. Is this a preview of a feature that isn't yet available in the first beta of Xcode 27? Is this because I'm using a local model or because it only works with a specific kind of model?

Hello, I’m trying to use a third-party MCP server (ICM) in Xcode 27’s new Coding Assistant. The server appears to start correctly and shows up when I run /mcp, but as soon as an agent tries to use it, Xcode appears to rewrite my config and changes "enabled = true" to "enabled = false". My configuration looks like this: [mcp_servers.icm] command = "/opt/homebrew/bin/icm" args = [ "serve", "--no-embeddings", "--db", "/Users/Sam/emdash/worktrees/Castle/emdash/custom-widget-system-5dubi/.icm-xcode/memories.db" ] enabled = true I’ve spent quite a while debugging this and found some interesting behaviour: The MCP server appears when I type /mcp. SQLite databases can be created and written inside the current workspace. SQLite databases can also be created and written in /tmp. SQLite databases outside the workspace (for example in ~/Library/Application Support) fail with readonly/open errors. Ordinary file operations outside the workspace still work fine. What I’m trying to understand is: Why is Xcode disabling the MCP server? What filesystem access should MCP servers have? Are they intentionally restricted to the workspace and /tmp? Is there a supported way to allow an MCP server to access data stored elsewhere on disk? Has anyone else run into this? Thanks!

Could you please add support for https://agents.md in Xcode coding intelligence? By default or at least optionally add the AGENTS.md file to coding agent sessions.

How to select GPT-5.5 in Chat? In Xcode 27 beta under Codex Settings I see GPT 5.5. But in "New Conversation" -> Chat / OpenAI there's only the old GPT 5 and 4.1!?

Hello there everyone. I've got numerous support projects using old school development. Things from Objective-C to RxSwift and UIKit. I've tried the LLM/Agents and they just destroy the codebase over time, so I have given up. The final straw was trying to get it to fix a POSTCODE validation issue which was using some novel ways to provide validation in place using RxSwift. Has anyone had any luck with these LLMs on Legacy codebases? Would love to hear tips.

We would like to use Xcode's new agentic coding feature (probably with Claude Agent). I had a look at the privacy statement, and under "Third-Party Agentic Coding Tools" it says the following: Capabilities include but are not limited to code searching, retrieving build logs, building source code, and arbitrary code execution. This is all expected, except for the bit about "arbitrary code execution". What does this mean, exactly? This sounds like a backdoor to my machine, surely that's not it? Does "code" refer to the code in my project, or any code the agent may provide? I know that other agentic coding tools may execute terminal commands, which you have to (or can select to) confirm first. Is that what is meant by arbitrary code execution? How arbitrary are we talking? And can I configure Xcode to always ask my permission first before any code is executed?

I enter the following example code into the chat screen, the code is applied, but the Undo button does not appear. The Undo button is also grayed out in the Edit menu. I don't know if this is due to the update or a settings issue, as this is my first time using Xcode. Before the update, there were no problems; the Undo button would appear right next to the chat screen after the code was applied, but now it's gone. Can you help me with this? Edit only: MetalPortfolio/MetalPortfolio/Models/Transaction.swift Goal: Make Transaction decoding backward-compatible so older saved transactions do not fail to load when newer fields are missing. Critical rules: - Do not edit any other file. - Do not change Transaction stored property names. - Do not change encoding format. - Do not change UI. - Do not change storage/export/import logic yet. - Do not change calculations. - Keep existing behavior for valid current JSON. Implementation: In Transaction.init(from:), change only optional/backward-compatible decoding where safe. Required: - fee should decode with default 0 if missing: fee = try container.decodeIfPresent(Double.self, forKey: .fee) ?? 0 Optional if consistent with existing defaults: - unit can keep decodeIfPresent fallback. - currency can keep decodeIfPresent fallback. Do not make core required fields optional: - id - type - asset - quantity - price - date Return only: - which file changed - exact decoding behavior changed

I've been investigating a persistent 401 Invalid Bearer Token error with Claude Agent in Xcode 26.3 on an Intel Core i9 Mac and wanted to share my findings here in case others are affected. ENVIRONMENT Hardware: Intel Core i9 (x86_64) Xcode: 26.3 (17C529) macOS: 26.3 (25D125) THE ISSUE Claude Agent fails with "Failed to authenticate. API Error: 401 - Invalid bearer token" on every prompt, despite the account showing as Signed In under Settings → Intelligence. The error persists after signing out and back in, and the token is confirmed valid and unexpired in Keychain. WHAT WORKS Claude Code in Terminal works perfectly on the same machine with the same credentials The basic Claude Sonnet 4.5 coding assistant in Xcode works fine Only Claude Agent is affected ROOT CAUSE HYPOTHESIS Xcode appears to be installing an ARM64 Claude binary on Intel Macs silently, with no warning or fallback. This binary likely fails to execute on x86_64 hardware and that failure is being misreported upstream as a 401 authentication error — which is why the error has nothing to do with the actual credentials. APPLE'S RESPONSE Feedback report FB22141224 was closed as 'Works as currently designed' with the explanation that Claude Agent requires the Neural Engine found on Apple Silicon. However, Claude Agent calls Anthropic's remote API over the network and performs no on-device AI processing — the Neural Engine is not involved. The fact that Claude Code works fine on this same Intel Mac with the same credentials demonstrates this clearly. WORKAROUND Using an Anthropic API key instead of OAuth resolves the issue but requires additional paid billing outside of an existing Claude.ai subscription. Has anyone else experienced this on either Intel or Apple Silicon? I'd be particularly interested to hear if Apple Silicon users are also affected, as the Developer Forums suggest this may not be limited to Intel Macs (thread/816369). Any input from Apple DTS engineers would be greatly appreciated.

Has anyone got Xcode working with Opus 4.7? I have Claude Code running 4.7, but Xcode insists on running 4.6. Even working with Claude to try to configure Xcode doesn't work. Apparently Xcode does something that prevents using Opus 4.7. Any reason for this? Xcode has a really nice integration, but I guess It's bak to Terminal and Claude Code if I want to use the better Model. Anyone else got this working?

Just installed the new Xcode 26.4 RC build (17E192) after happily using 26.3 for a few months. I'm noticing some immediate regressions in the Intelligence features: Frequent losing of OAuth token (Claude Agent). This had previously been fixed and now is back. Agent "Thinking" is constrained to thought bubble windows, which are (a) too small to read, (b) not scrollable when thinking goes beyond a few paragraphs. Yes they are tappable when thinking is finished, but this doesn't help. Due to (2), in deep thinking, it's impossible to follow progress beyond the visible window, so impossible to know if the agent is going off the rails. I'm noticing just more general slowness to complete tasks. Not just complete them -- it seems like it takes longer to START tasks, which is really weird. It sits there thinking for longer. Same project, same model as before. Every time you tap "New Chat" it presents both Claude and Codex choices, even if you're only signed into one. This turned a simple single tap into now a required two taps. Overall this feels like a frustrating setback after a very positive user experience in 26.3.

Codex will no longer be available in Xcode 26.3 starting on Friday, May 8th. Please upgrade to Xcode 26.4 or later to continue using Codex.

My Claude Agent component version was 2.1.81, it prompted me for an update the other day from the original release version 2.1.59. But now that I've installed Xcode 26.4.1 it's back to thinking 2.1.59 is the newest version. With every Xcode release it feels like there's a new regression or weird surprise with coding intelligence (26.4 was Claude losing its OAuth token every 8 hours or so). It really feels like this stuff needs more eyes on it. Filed as FB22529202

Since 26.4 beta 3 I've been having errors after a few hours of using Claude Agent. I get an error stating that I am not authed. If I go into Xcode settings, it still shows I am signed into Claude Agent. I have click sign out and then go through the whole auth process again to start using the agent again. This is on two separate computers. Is there anything I can do to stop this behaviour and stay logged in for longer? It worked fine in 26.3.

Hi, I'm running Xcode version 26.0.1 (17A400). Every time I restart Xcode, it seems to not recognize my Claude login anymore, and I have to go through the whole authentication flow again. It's quite annoying. Does anyone have a solution for it?

今天在Xcode上使用codeX，直接给我报错了，报错原因是这个： Codex encountered an error: {"detail":"The 'gpt-5.2-codex' model is not supported when using Codex with a ChatGPT account."}

Whenever I try and talk to chatgpt codex in Xcode, I get the request could not be completed error with this tag: JSON-RPC global stream failed: The operation couldn’t be completed. (IDEIntelligenceAgents.JSONRPCElement.Error error 1.) Chatgpt chat still works inside of xcode, however codex does not work. I have not hit my limit for codex. This issue happened after the xcode update. (I belive 26.4.1)

Environment: Xcode 26.3 RC macOS 26.3 (25D125) (Apple Silicon / arm64) Setup: Xcode > Settings > Intelligence > Claude Agent: Signed In (account status shows "Signed In") Model: Default Steps to reproduce: Open a new chat in Xcode's Coding Assistant Select "Claude Agent" from the agent dropdown (instead of "Claude Sonnet 4.5") Send any message (e.g. "HI") Expected result: Claude Agent responds normally. Actual result: The message is sent but immediately returns the error: "Your request couldn't be completed." 🚩I have an active Claude Code subscription with remaining usage. Running claude in Terminal works perfectly, confirming the subscription and quota are valid.

Setup: macOS 26.4 Xcode 26.4 Claude premium subscription Claude account logged in Xcode settings The issue usually surfaces after initially setting up Claude integration in Xcode and then leaving Xcode open for at least a day or so. The integration works initially (right after logging in). But then after some time it starts responding with this error. Failed to authenticate. API Error: 401 {"type":"error","error":{"type":"authentication_error","message":"OAuth token has expired. Please obtain a new token or refresh your existing token."},"request_id":"req_011Ya....."} Restarting Xcode does not help. The only thing that helps is log out from Claude account and log back in. Sounds like the token refresh logic is broken or missing. Please improve moving forward.

Hi guys, Two annoying issues I'm having at the moment with agentic coding. Xcode signs out of Claude every single day and gives the error Failed to authenticate. API Error: 401. The only solution is go to settings and sign out of Claude and in again. Apple removed the little arrow to select the model, so now every time you start a new conversation you have to pick the model. I just want to start a new conversation with the model I'm using. The extra click is a time waster. Any solutions would be appreciated. Thanks