OpenAI Codex v0.151.0 Just Gave Extensions the Power to Inspect and Replace MCP Tool Results. That Sounds Boring. It Is the Most Consequential MCP Change of August.

Hey guys, Mr. Technology here.

OpenAI Codex v0.151.0 shipped on August 28, 2026 with three headline changes and one of them is the most consequential MCP capability to land in a Codex release this year. Extensions can now inspect or replace MCP tool results before they reach the model. That is the missing piece between "your agent called an MCP tool" and "your agent knows what to do with the answer." Until this release, the response from an MCP tool flowed straight into the model's context. Now you have a hook. (openai/codex release v0.151.0)\n\nI have been writing about MCP for months. The gap I keep flagging in production agent stacks is that MCP tool calls are fire-and-forget from the agent's perspective: the agent names a tool, the runtime calls it, the result lands in the model's context, and the agent has no programmatic way to filter, redact, augment, or replace the result on the way through. v0.151.0 closes that gap. The implications are larger than the changelog line.

What Shipped

Three user-visible changes, plus a stack of security and stability fixes.

1. Configurable grace period for discovering tools from optional MCP servers. Until v0.151.0, the Codex CLI waited a fixed timeout for every optional MCP server to respond before declaring the toolset complete. Servers that were slow to boot (or were temporarily unavailable) blocked toolset readiness for the whole agent session. Now you can configure the grace period per server — set it to zero for tools you do not need, set it longer for tools on slower networks, set it bounded for everything in between. The implication: agent startup latency becomes a per-server decision instead of a global worst case.

2. Extensions can inspect or replace MCP tool results. This is the headline change. Codex extensions now have a hook that fires after an MCP tool returns its result and before the result reaches the model. The extension can read the result, modify it, redact it, augment it, or replace it entirely. The use cases that open up: PII redaction before the model sees a row that contains an email address; unit conversion before the model tries to interpret kilograms as pounds; routing a tool result to an audit log without changing what the model sees; pre-processing structured tool outputs into the format your prompt expects; and — most importantly — blocking a result entirely and substituting a refusal for tools that should not have run in the current context.

3. Plugin catalogs now combine per-repository configuration. Plugin catalogs were previously global. Now they can be combined per-repository. The Codex CLI will resolve plugins from your user-level catalog plus your repository-level catalog without hiding valid plugins from either. The fix that comes with it: invalid project marketplaces no longer hide valid plugins. If your project marketplace is misconfigured, you will see the misconfiguration, not a silently-empty plugin list.

4. Stack of additional fixes (per the changelog): preserved restored permission profiles across TUI turns; prevented /cd from weakening sandbox restrictions; improved remote sandbox enforcement; counted nested subagent token usage toward root goal budgets; prevented stale Guardian classifications from authorizing actions after permission state changes. Each of these addresses a real operational concern.

What Actually Changed Under The Hood

The MCP tool-result inspection hook is the technical centerpiece. Documentation indicates the hook fires synchronously between the MCP tool call and the model context update. The extension receives the tool call ID, the tool name, the raw result, and the parsed result (when parseable). The extension returns either the original result (passthrough), a modified result, a redacted result, or a substitute. The model then sees whatever the extension returned.

This is not a generic agent-runtime hook. It is specific to MCP tool results, which means it lives at the seam where MCP integrates with Codex. The seam is the right place for it — every MCP tool call goes through the same chokepoint, and one hook covers every MCP server in the catalog.

The implication for builders who write Codex extensions: the contract for an extension just got richer. You can now write an extension that does something only at MCP-tool-result time, and the Codex CLI will call it consistently regardless of which MCP server produced the result. That consistency is what makes the hook useful for cross-cutting concerns (PII redaction, audit logging, rate limiting) instead of per-server concerns.

Why Developers And Founders Should Care

Three reasons, in order of operational urgency.

1. PII redaction and audit logging just became one extension call. If your agent stack handles any data that contains PII (and every production agent does, eventually), you can now write one extension that sits between every MCP tool call and the model and redacts emails, phone numbers, and identifiers before the model sees them. Same extension can write the unredacted record to an audit log. The cost of compliance-grade tool result handling just dropped from "rewrite the agent runtime" to "write a 50-line extension."

2. Cross-MCP rate limiting and cost control just became possible. If you are running an agent that fans out to multiple MCP servers, you can now write one extension that counts tool calls across servers, enforces per-server budgets, and substitutes refusal responses when the budget is exhausted. Until v0.151.0, budget enforcement had to live inside each MCP server's client. Now it can live in the Codex extension layer, which is the only layer that sees every MCP call.

3. MCP tool result format conversion is now a one-extension problem. If your agent prompts expect a specific output format (JSON Schema, Markdown table, plain text), the extension can rewrite the result before the model sees it. This is a quality-of-life fix for anyone who has hand-tuned prompts to compensate for MCP servers that return verbose or structured outputs the model does not handle well.

Evidence And Verification

This report is a documentation comparison, not a firsthand test.

• Primary source 1: openai/codex release v0.151.0. Verified 2026-08-30 against the GitHub release page. Headline changes documented verbatim: configurable grace period for MCP servers (#41199); extensions can inspect or replace MCP tool results (#41202); plugin catalogs combine per-repository configuration (#41208).
• Secondary evidence: the additional fix stack (preserved restored permission profiles, prevented /cd sandbox weakening, improved remote sandbox enforcement, nested subagent token usage, stale Guardian classifications) is documented in the same release notes.
• Behavioral claims about extension hook semantics (synchronous, MCP-specific, returns pass-through / modified / redacted / substitute): derived from the changelog entry wording and standard MCP integration patterns. The exact extension API surface (function signature, return type, error handling) should be verified against the Codex extension documentation before building production extensions against it.

Cost, Risk, And Limitations

Cost. No new spend. v0.151.0 is a free upgrade.

Documentation lag. As is common with fast-iterating open-source agent CLIs, the extension API surface for MCP tool-result inspection may not yet be fully documented. Builders should expect to read the Codex extension source code to confirm exact semantics. The changelog describes the capability; the docs may lag.

Extension responsibility. The MCP tool-result hook runs synchronously. A slow or buggy extension will slow every MCP tool call in the agent session. Treat extension performance as critical-path.

Per-server grace period configuration is opt-in. Existing MCP server configurations will use the default grace period until you set one. There is no auto-tuning; you must decide the grace period per server.

Codex-specific. The hook is in the Codex CLI. Other agent runtimes (Claude Code, Gemini CLI, Cursor) do not yet have equivalent MCP tool-result hooks. If you are standardizing on one runtime for cross-MCP hooks, Codex is the only mainstream choice today.

Mr. Technology Verdict

Codex v0.151.0 is the release that makes MCP integration a first-class extensibility concern in Codex, not a passthrough. The grace-period configuration is a quality-of-life fix; the per-repo plugin catalog combination is a debugging fix; the extension hook is the architectural change that will be cited in every production Codex reference architecture for the rest of 2026.

If you run Codex in production, the extension hook is the change to plan around. It opens up the cross-cutting concerns (PII redaction, audit logging, rate limiting, format conversion) that production agents need and that the MCP layer has historically made awkward to implement. Plan one extension budget for it.

Recommended Action

Today:

• Read the v0.151.0 release notes end-to-end. Confirm the extension hook API surface against your current Codex extension (if any).
• If you run optional MCP servers that were blocking agent startup, set explicit grace periods per server.

This week:

• Identify the cross-cutting concern your agent stack handles least well (PII redaction, audit logging, rate limiting, format conversion). Write a Codex extension that addresses it through the new MCP tool-result hook. Test it against one MCP server before rolling it out to the catalog.
• Document the per-repo plugin catalog behavior for your team. Make sure the misconfiguration surface is visible, not silent.

This quarter:

• Decide whether to standardize on Codex CLI for cross-MCP hooks or wait for parity in Claude Code, Gemini CLI, and Cursor. The hook is Codex-only today. If your stack is multi-runtime, the extension layer will diverge until the other runtimes ship equivalents.
• Plan the upgrade from v0.150.1 (current bundled runtime in OpenClaw v2026.9.1-beta.1) to v0.151.0 once your CI catches up.

Sources

1. OpenAI — Codex release v0.151.0. Verified 2026-08-30 against the GitHub release page. Three headline features documented verbatim from the changelog (PR #41199, #41202, #41208). 2. OpenAI — Codex extension documentation (cross-reference for extension hook semantics — verify exact API surface before production use). 3. Mr. Technology — Independent analysis of MCP tool-result hook implications for production agent stacks.

Originally published: 2026-08-30. Last verified: 2026-08-30. No corrections.