OpenAI Codex release notes: alpha adds audio and classifies multi-agent v2 as stable

The stable Codex CLI release is still 0.144.6. The new work is in pre-release 0.145.0-alpha.28, published on July 21, 2026. Dynamic tools and code mode can now pass audio to the model. At the same time, multi_agent_v2 is classified as stable in the codebase, though it remains disabled by default.

Codex CLI is OpenAI's local coding agent for the terminal. A dynamic tool can produce a response that the agent uses as new input. MCP, or Model Context Protocol, is a protocol for connecting AI agents to tools and data sources. A multi-agent workflow divides work among several coordinated agent roles.

OpenAI Codex release notes: two signals in the 0.145 alpha

OpenAI's GitHub release labels 0.145.0-alpha.28 as a pre-release. The npm registry points to alpha.28 as the current alpha while latest remains 0.144.6. This is a test track, not a stable upgrade recommendation.

Human step: Use a separate alpha environment that is already intended for pre-release testing and confirm it is running 0.145.0-alpha.28. OpenAI's sources do not document a public activation procedure for multi_agent_v2, so this article does not invent a command or menu path.

Sources: OpenAI Codex 0.145.0-alpha.28 on GitHub. npm registry for @openai/codex.

Audio can now travel from a tool to the model

The change adds inputAudio content items to dynamic tool responses, app-server events, thread history, and generated protocol schemas. Code mode also gets an audio() helper that accepts inline data URLs, audio URL objects, and MCP audio blocks.

Two limits matter. Ordinary remote audio URLs are rejected, and a model without audio support receives an explanatory text item instead. The release does not mean Codex can listen to any audio link. It opens a testable route from an audio-producing tool to an audio-capable model.

Source: PR #34080, audio support for dynamic tools and code mode.

Multi-agent v2 leaves the development warning but does not switch itself on

OpenAI now classifies multi_agent_v2 as stable in the Codex codebase. Structured configuration for the feature should no longer trigger the under-development warning. The same change explicitly keeps the feature disabled by default.

For teams that want to divide a run between a lead agent and an independent reviewer, the classification matters. The source does not claim general availability or document a public activation command. Production assumptions should wait for that documentation.

Source: PR #34383, multi-agent v2 classified as stable.

Short example: use the new Codex feature

Run this prompt in an alpha test environment after a human has selected the right version and opened the right codebase:

Find dynamic tools or MCP tools in this codebase that can return audio. Trace the path from tool response to model input and thread history. Then propose one independent reviewer role for a multi-agent v2 test. Cite files and tests, flag remote-audio assumptions, and change nothing.

Good output should:

• Show where audio is created, validated, and retained in history.
• Distinguish inline data and MCP audio blocks from ordinary remote URLs.
• Give the reviewer clear inputs, a stop point, and a deliverable.
• Leave the codebase unchanged.

What teams can test now

Choose an existing workflow that already handles recordings, customer calls, or spoken work notes. Check whether audio can move through the tool chain without being reduced to a file reference the model cannot interpret. Then have a separate reviewer role check source citations or transcription results.

When the trial connects to real systems, Tool Forge can help with scoped tool permissions, secrets in environment variables or a secret manager, approval before writes, and retained run logs.