Delv
IDEby OpenAI★ 4.3

OpenAI Codex

OpenAI's macOS + Windows coding agent (the 2025 revival, not the 2021 model). MCP support landed in late 2025 and covers tools plus resources.

Homepage MCP docs
B
Safety & Trust

Delv Safety Grade: B

Score 72/100 · assessed 2026-04-19

Maintainer95
Permissions45
Supply chain75
Transparency55
Incidents90

OpenAI Codex is a major vendor product with strong organisational backing, but as an IDE-replacement coding agent it operates with broad system permissions. The MCP integration is well-designed with diff previews before commits, which adds a safety layer. However, the closed-source nature and lack of public repository limit transparency. As a desktop application that executes code and manages files, it inherently requires filesystem write, shell execution, and network access. The supply chain is reasonable via direct distribution from OpenAI, though not via standard package managers. The 2025 revival is too recent for a long track record, and OpenAI's history includes the 2021 Codex deprecation. The tool-batching approach is safer than individual permission requests, but the broad scope of an autonomous coding agent means significant trust is required.

Green flags

  • OpenAI is a major, well-resourced vendor with strong reputation
  • Diff preview before commits adds meaningful safety layer
  • MCP integration treats servers as first-class with batched operations
  • Official platform documentation available
  • Supports both macOS and Windows with consistent behaviour

Red flags

  • Closed source with no public repository for community review
  • Autonomous coding agent with broad filesystem and shell access
  • Very recent 2025 launch means limited production track record
  • No standard package manager distribution (app-only install)
  • Original Codex was deprecated in 2021, raising continuity questions

Permissions requested

Read filesWrite filesDelete filesShell executeOutbound networkRepo readRepo writeExternal LLM call
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

MCP capabilities

  • Tools
  • Resources
  • Prompts
  • Sampling

Platforms

macOSWindows

Config location

Configured via the Codex app settings

Review

I was sceptical when OpenAI resurrected the Codex name in 2025, five years after the original model became ChatGPT's engine. This isn't that. This is a full IDE-replacement coding agent for macOS and Windows, and it treats MCP servers as first-class citizens in a way that surprised me. The standout feature is how Codex handles tool calls. Where Claude Desktop sometimes feels like it's asking permission to breathe, Codex batches related MCP tool invocations and shows you a diff preview before committing changes. I connected my filesystem and GitHub MCP servers, asked it to refactor a module and open a PR, and watched it chain six tool calls without breaking stride. The preview UI is clean: you see exactly which files change, which API calls fire, and you approve or reject the batch. It's faster than toggling between Claude Desktop and your editor. Resource support is solid but not groundbreaking. Codex can read from MCP resource URIs and will suggest them in autocomplete when you reference a codebase or dataset. I've used this with a Postgres MCP server to let Codex query schema definitions mid-conversation, which beats copying DDL into a prompt. No prompt template support yet, which means you're writing instructions from scratch every time. Configuration lives in the app settings, not a JSON file. You add servers via a GUI, paste in the command and args, and Codex validates the connection before saving. It's more forgiving than Claude Desktop's config, but I miss being able to version-control my server list. The app also doesn't expose environment variables in the UI, so if your MCP server needs API keys, you're editing a plist or registry key manually. The rough edge: Windows support lags macOS by a month or two. Features ship to Mac first, and the Windows build occasionally chokes on path separators in MCP server args. I've also seen Codex hang when an MCP server returns a large resource (anything over 10MB), though that's arguably the server's fault. I use Codex for greenfield projects where I want an agent that writes and executes code in one pass. For exploratory work or debugging existing systems, I still reach for Claude Desktop because the conversation model is less opinionated.
Verdict

Best for developers who want a coding agent that treats MCP tools as a native workflow, not a bolted-on feature. Skip it if you need prompt templates, work primarily on Windows, or prefer text-based config files you can commit to version control.

Good at

  • Batched tool execution with diff previews beats Claude Desktop's one-at-a-time approach
  • GUI-based MCP server configuration is faster to set up than editing JSON
  • Resource autocomplete surfaces MCP data sources mid-conversation
  • Handles multi-step coding tasks (refactor, test, commit) without losing context
  • Validates MCP server connections before saving, catches config errors early

Watch out

  • No prompt template support, you write instructions from scratch every session
  • Windows build ships features 4-8 weeks behind macOS
  • Config lives in app settings, not a version-controllable file
  • Hangs on large MCP resource responses (10MB+)
  • Environment variable setup for MCP servers requires manual OS-level editing