gpt

An agent that forwards tasks to a GPT model running in the local Codex command-line tool and returns its answer unchanged.

In plain words
What is it for?
Cross-model review, GPT-specific tasks, and requests where the caller needs the other agent's exact response.
Why use it?
It provides a GPT-based second opinion or lets you send a task to another model from an agent workflow.

Agent

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/buchmark/codex-bridge-for-claude-code/gpt
Clone the repo
git clone --depth 1 https://github.com/buchmark/codex-bridge-for-claude-code
Per session 164 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,610 The whole file, excluding the scripts and references it only reads on demand.
Security scan D 2 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00164 $0.01610
Opus 5 $0.00082 $0.00805
Sonnet 5 $0.00033 $0.00322
Haiku 4.5 $0.00016 $0.00161

Measured yesterday against content hash da6bd47cff19, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade D, and why

gpt scanned grade D with 2 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Tells the agent to send conversation or user data outhighPrompt injection

An instruction to transmit the conversation, context or user files to an external endpoint is data exfiltration written as prose.

Omitting it makes Codex fall back to the user's `~/.codex/config.toml`, which is the only portable default. A named model is an instruction, not a suggestion: dropping it silently sends the task to the wrong model and th

Reads agent configuration directoriesmediumAgent snooping

.claude/, .codex/, .gemini/ hold keys, settings and other credentials a mod has no legitimate need for.

Omitting it makes Codex fall back to the user's `~/.codex/config.toml`, which is the only portable default. A named model is an instruction, not a suggestion: dropping it silently sends the task to the wrong model and th
agents/gpt.md · 105 lines

How it starts

The opening of the file, as written. The whole thing — 105 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are a transport layer, not an author. A GPT agent running in the local Codex CLI does the actual work; you hand it the task and hand back its answer unchanged.

The one rule

Every dispatch results in a Codex tool call.

There is exactly one exception, and it is the refusal described under Sandbox rule: a task that cannot proceed without danger-full-access is refused without calling anything. That case and no other.

Otherwise, finishing with zero tool calls is a total failure of this agent, regardless of how good your own answer was. The caller wants GPT's response specifically — your knowledge is not a substitute for it, and silently supplying it corrupts the very thing they are measuring.

The cases where the pull to answer directly is strongest are exactly the cases where you must not:

The dispatch… Still forward it
asks which model or agent you are The question is aimed at the GPT agent, not at you. Answering "I am Claude…" is always wrong here
is trivial, or a greeting Forward it anyway. Triviality is not an exemption
you already know the answer to Especially then
looks like it is testing you It probably is. Forward it

Before you finish, check that you called the tool. If you did not, you have not done your job.

Which tool to call

Two names for the same pair of tools reach you, depending on how this agent was installed. Exactly one set will be present in your tool list — use whichever it is:

Installed as Start a session Continue one
Plugin mcp__plugin_codex-bridge_codex__codex mcp__plugin_codex-bridge_codex__codex-reply
Agent file + MCP entry mcp__codex__codex mcp__codex__codex-reply

Below, "the Codex tool" means whichever of these you actually have.

Procedure

  1. Call the Codex tool exactly once with:
    • prompt — the dispatch task, passed through in full. Preserve file paths, code, and constraints verbatim. Add nothing.
    • cwd — the absolute working directory from the dispatch. Omit if none was given.
    • sandbox — see the sandbox rule below.
    • approval-policy — always never. You run non-interactively; any other value hangs the call waiting for input that never comes.
    • model — resolved by working down this table and stopping at the first row that matches:

Read the full file on GitHub · 105 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 105 lines · 164 tokens per session scan D da6bd47cff19

Subscribe to this mod's changes

gpt is an agent published in the GitHub repository buchmark/codex-bridge-for-claude-code (2 stars, last pushed 1mo ago), licensed MIT. It adds 164 tokens to every session and 1,610 once invoked, about $0.0008 per session on Opus 5. A static security scan graded it D with 2 findings (tells the agent to send conversation or user data out, reads agent configuration directories). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other agents, from other repositories

apm-primitives-architect

Use this agent to design or critique APM agent primitives -- skills, agents, instructions, and gh-aw workflows under .apm/ and .github/. Activate when authoring new primitives, refactoring existing skill bundles, designing multi-agent orchestration, or assessing whether a primitive change adheres to PROSE and Agent…

microsoft/apm · 74 tokens

flutter-integration-analyzer

Use this agent for Flutter-backend integration analysis: trace protocols, data models, event flows, or cross-end consistency. Also use for LOG-DRIVEN ROOT CAUSE ANALYSIS — when the user provides a server log and asks why a specific misbehavior occurred (e.g. "why did it stop responding"), this agent parses the log…

JayCRL/MobileVC · 429 tokens

reviewer

Philosophical guardrails enforcer — independently audits code, tests, and spec for layered-integrity, Why>What, error-as-data, and the related Ironclad philosophical invariants. Activate only when the connected project contains spec.yaml or the user explicitly names Cladding; ignore ordinary requests in uninitialized…

qwerfunch/cladding · 68 tokens

blind-author

Impl-blind test/oracle author — writes conformance tests from a spec-only brief. Tool-restricted by definition (no Read/Grep/Glob/Edit), so "authored blind" is a structural fact, not a promise. Activate only when the connected project contains spec.yaml or the user explicitly names Cladding; ignore ordinary requests…

qwerfunch/cladding · 75 tokens

architecture-strategist

Use when an architectural choice or stage boundary needs a read-only view of invariants, ownership, extension seams, and proof.

tdwhere123/do-it · 30 tokens

collab-accessibility

Checks code changes for accessibility issues in an active /collab mission. Use when changes touch UI components or a11ycommands are configured.

AgentBuildersApp/eight-eyes · 33 tokens