Add MCP Sampling Support to Enable Agentic MCP Tools #5116
Conversation
- Change CodexSamplingHandler to store Config instead of ModelClient - Create independent LLM calls for each sampling request - Use sampling request's systemPrompt (or empty if not provided) - Select model from model_preferences.hints if available - Skip system message in chat completions when instructions are empty Per MCP spec, sampling requests should use their own systemPrompt and model preferences, not inherit from the session's configuration. Fixes the issue where MCP sampling was using Codex's full agent instructions instead of the server's systemPrompt. Might close openai#4929
|
All contributors have signed the CLA ✍️ ✅ |
yaonyan
changed the title
|
I have read the CLA Document and I hereby sign the CLA |
yaonyan
changed the title
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
|
An example MCP server (with sampling request) is required to test out the sampling feature. I will provide one for reference: experimental_use_rmcp_client = true
[mcp_servers.code-reviewer]
args = [
"-y",
"@mcpc-tech/cli",
"--config-url",
"https://raw.githubusercontent.com/mcpc-tech/mcpc/refs/heads/main/packages/cli/examples/configs/code-reviewer.json",
]
command = "npx"
env = { "MCPC_TRACING_ENABLED" = "true" }
startup_timeout_sec = 36
tool_timeout_sec = 3600 |
…s based on RMCP client usage
|
@gpeal, would you mind taking a look here? This would make codex the first CLI to support MCP sampling. |
|
Thanks for your work here. Unfortunately, this isn't something we can accept at this time. Although there are many valid use cases for this, we don't yet have an appropriate way to prevent people from abusing this to turn their ChatGPT Codex credits into generic completions API credits. We will continue to think through alternatives here and may revisit this in the future. |
3 participants
Summary
Implements MCP
sampling/createMessagecapability, allowing MCP servers to make independent LLM calls through Codex with their own prompts and model preferences.Closes #4929, follows mcp sampling spec: https://modelcontextprotocol.io/specification/2025-06-18/client/sampling
Key Changes
New
CodexSamplingHandler(codex-rs/core/src/mcp_sampling_handler.rs)systemPromptandmodel_preferences.hintsIntegration:
chat_completions.rs: Skip system message when instructions are emptyrmcp-client: AddedSamplingHandlertrait for delegationMcpConnectionManager: Wires sampling handler and declares capabilityBenefits
Write mini agents inside Codex using MCP tools. See this code review agent demo below - it can also run in background.
Checks Passed
✅ cargo test -p codex-core
✅ just fix -p codex-core
✅ just fix -p codex-rmcp-client