OpenAI Codex release notes: 0.147 alpha tests MCP 2026 and ordered threads

OpenAI published Codex CLI 0.147.0-alpha.1 on July 29 and 0.147.0-alpha.2 on July 30, 2026. The two test releases contain two changes worth tracking: Codex can test the July 28 MCP revision, and clients get better building blocks for organizing many parallel threads. The stable release is still 0.146.0, so this is a testing opportunity rather than a general production upgrade.

OpenAI Codex release notes: what is in the 0.147 alpha?

Codex CLI is OpenAI's local coding agent for the terminal. A coding agent can read a repository, propose changes, and use tools within the permissions granted by its user.

The GitHub pages for alpha 1 and alpha 2 contain only short release lines. The practical detail comes from the pull requests included in those tags. npm currently lists 0.146.0 as latest and 0.147.0-alpha.2 as alpha.

Source: OpenAI Codex 0.147.0-alpha.1.

Source: OpenAI Codex 0.147.0-alpha.2.

Source: npm registry metadata for @openai/codex.

MCP 2026-07-28 becomes an opt-in test mode

MCP, the Model Context Protocol, is an open protocol that connects AI applications to external tools and data sources. Revision 2026-07-28 uses a stateless core in which each request carries version and capability data. server/discover lets a client ask which protocol versions a server supports.

The Codex alpha adds the mcp_2026_07_28 feature flag. It is disabled by default and marked under development in the tagged source. When enabled, Codex can try server/discover over streamable HTTP, consume paginated tool and resource catalogs, and fall back when the response establishes that the server supports only the legacy protocol. Stdio servers also require CODEX_MCP_PROTOCOL_VERSION=2026-07-28 in the server configuration.

The second part matters more than the version number: tools/call and resources/read can continue across multiple rounds through input_required. An integration can ask for missing information without discarding the state of the current tool request. The flow covers JSON, SSE, and stdio transports.

Source: OpenAI PR #35724 on MCP discovery and compatibility.

Source: OpenAI PR #35725 on multi-round tools and input_required.

Source: MCP specification revision 2026-07-28.

Alpha 2 also lets unrelated tools continue while an MCP server starts. It reports authentication as unknown when OAuth discovery is inconclusive instead of incorrectly calling support unsupported. That gives operations teams a better diagnostic: a transient network failure no longer looks like a definitive capability result.

Source: OpenAI PR #35937 on tool execution during MCP startup.

Source: OpenAI PR #36045 on unknown MCP authentication status.

Threads gain sections, ordering, and named forks

Stable 0.146.0 added names for new sessions and support for pinned threads. The 0.147 alpha builds on that foundation. The app server replaces a simple isPinned field with persisted sections, including a stable section for pinned threads. Clients can list sections even when they are empty and filter threads by section.

Alpha 2 adds thread/section/move for moving and ordering threads within a section. The order can be stored in SQLite and survive a restart. The terminal UI also accepts an optional name after /fork, such as /fork oauth-fallback, and persists the name on the new branch.

These are primarily app-server and alpha capabilities. The pull requests do not promise a complete sections interface in every Codex client. /fork <name> is explicitly verified for the terminal UI in alpha 2.

Source: OpenAI PR #35722 on persisted thread sections.

Source: OpenAI PR #36007 on manual section ordering.

Source: OpenAI PR #36036 on names after /fork.

What this changes for Nordic teams

For a Nordic integration team, the test is concrete: keep the work organized in threads while the MCP connection handles modern discovery, multi-round requests, and clearer status. This fits workflows in which Codex reads from an issue tracker or documentation system, prepares a proposal, and waits for a person before a writing tool runs.

Start with one bounded integration the team actually uses. Store access tokens in environment variables or a secret manager, give the server the minimum required permissions, and require approval for writing tools. Record discovery results, authentication status, and tool calls in a run receipt. The team can then judge the protocol without turning security into the entire project.

Human step

Use this only with a 0.147.0-alpha build in a test environment. OpenAI's configuration guide verifies the codex --enable feature_name syntax, and the tagged alpha source registers the mcp_2026_07_28 key.

Start Codex with the test mode:

codex --enable mcp_2026_07_28

For a stdio MCP server, also add the version marker to that server's existing config.toml entry:

[mcp_servers.your_server.env]
CODEX_MCP_PROTOCOL_VERSION = "2026-07-28"

Replace your_server with the name of the server you have already configured. HTTP servers do not use this stdio marker.

Source: OpenAI Codex feature flag guide.

Source: OpenAI Codex guide for MCP server env configuration.

Source: OpenAI PR #35724 with the exact stdio marker.

Short example: use the new Codex feature

Send this prompt after the test mode and selected MCP server are running:

Inspect the configured MCP integration and produce a readable compatibility receipt for the 2026-07-28 protocol. Report transport, discovery result, catalog pagination, any input_required rounds, authentication status, and legacy fallback. Run no write tools and change no files.

Good output should:

• Name the transport and protocol version that was actually discovered.
• Separate verified results from checks that could not be completed.
• Show whether authentication is supported, unknown, or failed.
• Leave the codebase and external systems unchanged.

What should you watch next?

Watch whether the feature flag leaves its "under development" stage, whether OpenAI documents broader client support for sections, and whether 0.147 becomes stable. Until then, 0.146.0 remains the normal npm release. Alpha 2 is most relevant to integration developers who can run a controlled compatibility test and inspect the logs.

This differs from the July 29 article about stable 0.146.0, named sessions, and Agent Plugins. Today's signal is the alpha branch's MCP 2026 support, multi-round tools, and thread sections.

When the test connection needs to become a reusable workflow, Tool Forge is the natural next step: document the server, permissions, approval points, and run receipt before connecting more systems.