## Summary
- Add the executor-backed RMCP stdio transport.
- Wire MCP stdio placement through the executor environment config.
- Cover local and executor-backed stdio paths with the existing MCP test
helpers.
## Stack
```text
o #18027 [6/6] Fail exec client operations after disconnect
│
@ #18212 [5/6] Wire executor-backed MCP stdio
│
o #18087 [4/6] Abstract MCP stdio server launching
│
o #18020 [3/6] Add pushed exec process events
│
o #18086 [2/6] Support piped stdin in exec process API
│
o #18085 [1/6] Add MCP server environment config
│
o main
```
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- honor `_meta["codex/imageDetail"] == "original"` on MCP image content
and map it to `detail: "original"` where supported
- strip that detail back out when the active model does not support
original-detail image inputs
- update code-mode `image(...)` to accept individual MCP image blocks
- teach `js_repl` / `codex.emitImage(...)` to preserve the same hint
from raw MCP image outputs
- document the new `_meta` contract and add generic RMCP-backed coverage
across protocol, core, code-mode, and js_repl paths
## Changes
Allows MCPs to opt in to receiving sandbox config info through `_meta`
on model-initiated tool calls. This lets MCPs adhere to the thread's
sandbox if they choose to.
## Details
- Adds the `codex/sandbox-state-meta` experimental MCP capability.
- Tracks whether each MCP server advertises that capability.
- When a server opts in, `codex-core` injects the current `SandboxState`
into model-initiated MCP tool-call request `_meta`.
## Verification
- added an integration test for the capability
## Why
For more advanced MCP usage, we want the model to be able to emit
parallel MCP tool calls and have Codex execute eligible ones
concurrently, instead of forcing all MCP calls through the serial block.
The main design choice was where to thread the config. I made this
server-level because parallel safety depends on the MCP server
implementation. Codex reads the flag from `mcp_servers`, threads the
opted-in server names into `ToolRouter`, and checks the parsed
`ToolPayload::Mcp { server, .. }` at execution time. That avoids relying
on model-visible tool names, which can be incomplete in
deferred/search-tool paths or ambiguous for similarly named
servers/tools.
## What was added
Added `supports_parallel_tool_calls` for MCP servers.
Before:
```toml
[mcp_servers.docs]
command = "docs-server"
```
After:
```toml
[mcp_servers.docs]
command = "docs-server"
supports_parallel_tool_calls = true
```
MCP calls remain serial by default. Only tools from opted-in servers are
eligible to run in parallel. Docs also now warn to enable this only when
the server’s tools are safe to run concurrently, especially around
shared state or read/write races.
## Testing
Tested with a local stdio MCP server exposing real delay tools. The
model/Responses side was mocked only to deterministically emit two MCP
calls in the same turn.
Each test called `query_with_delay` and `query_with_delay_2` with `{
"seconds": 25 }`.
| Build/config | Observed | Wall time |
| --- | --- | --- |
| main with flag enabled | serial | `58.79s` |
| PR with flag enabled | parallel | `31.73s` |
| PR without flag | serial | `56.70s` |
PR with flag enabled showed both tools start before either completed;
main and PR-without-flag completed the first delay before starting the
second.
Also added an integration test.
Additional checks:
- `cargo test -p codex-tools` passed
- `cargo test -p codex-core
mcp_parallel_support_uses_exact_payload_server` passed
- `git diff --check` passed
- Changed `requires_mcp_tool_approval` to apply MCP spec defaults when
annotations are missing.
- Unannotated tools now default to:
- `readOnlyHint = false`
- `destructiveHint = true`
- `openWorldHint = true`
- This means unannotated MCP tools now go through approval/ARC
monitoring instead of silently bypassing it.
- Explicitly read-only tools still skip approval unless they are also
explicitly marked destructive.
**Previous behavior**
Failed open for missing annotations, which was unsafe for custom MCP
tools that omitted or forgot annotations.
---------
Co-authored-by: colby-oai <228809017+colby-oai@users.noreply.github.com>
## What?
- Render an MCP image output cell whenever a decodable image block
exists in `CallToolResult.content` (including text-before-image or
malformed image before valid image).
## Why?
- Tool results that include caption text before the image currently drop
the image output cell.
- A malformed image block can also suppress later valid image output.
## How?
- Iterate `content` and return the first successfully decoded image
instead of only checking the first block.
- Add unit tests that cover text-before-image ordering and
invalid-image-before-valid.
## Before
```rust
let image = match result {
Ok(mcp_types::CallToolResult { content, .. }) => {
if let Some(mcp_types::ContentBlock::ImageContent(image)) = content.first() {
// decode image (fails -> None)
} else {
None
}
}
_ => None,
}?;
```
## After
```rust
let image = result
.as_ref()
.ok()?
.content
.iter()
.find_map(decode_mcp_image)?;
```
## Risk / Impact
- Low: only affects image cell creation for MCP tool results; no change
for non-image outputs.
## Tests
- [x] `just fmt`
- [x] `cargo test -p codex-tui`
- [x] Rerun after branch update (2026-01-27): `just fmt`, `cargo test -p
codex-tui`
Manual testing
# Manual testing: MCP image tool result rendering (Codex TUI)
# Build the rmcp stdio test server binary:
cd codex-rs
cargo build -p codex-rmcp-client --bin test_stdio_server
# Register the server as an MCP server (absolute path to the built binary):
codex mcp add mcpimg -- /Users/joshka/code/codex-pr-review/codex-rs/target/debug/test_stdio_server
# Then in Codex TUI, ask it to call:
- mcpimg.image_scenario({"scenario":"image_only"})
- mcpimg.image_scenario({"scenario":"text_then_image","caption":"Here is the image:"})
- mcpimg.image_scenario({"scenario":"invalid_base64_then_image"})
- mcpimg.image_scenario({"scenario":"invalid_image_bytes_then_image"})
- mcpimg.image_scenario({"scenario":"multiple_valid_images"})
- mcpimg.image_scenario({"scenario":"image_then_text","caption":"Here is the image:"})
- mcpimg.image_scenario({"scenario":"text_only","caption":"Here is the image:"})
# Expected:
# - You should see an extra history cell: "tool result (image output)" when the
# tool result contains at least one decodable image block (even if earlier
# blocks are text or invalid images).
Fixes#9814
---------
Co-authored-by: Josh McKinney <joshka@openai.com>
It's pretty amazing we have gotten here without the ability for the
model to see image content from MCP tool calls.
This PR builds off of 4391 and fixes#4819. I would like @KKcorps to get
adequete credit here but I also want to get this fix in ASAP so I gave
him a week to update it and haven't gotten a response so I'm going to
take it across the finish line.
This test highlights how absured the current situation is. I asked the
model to read this image using the Chrome MCP
<img width="2378" height="674" alt="image"
src="https://github.com/user-attachments/assets/9ef52608-72a2-4423-9f5e-7ae36b2b56e0"
/>
After this change, it correctly outputs:
> Captured the page: image dhows a dark terminal-style UI labeled
`OpenAI Codex (v0.0.0)` with prompt `model: gpt-5-codex medium` and
working directory `/codex/codex-rs`
(and more)
Before this change, it said:
> Took the full-page screenshot you asked for. It shows a long,
horizontally repeating pattern of stylized people in orange, light-blue,
and mustard clothing, holding hands in alternating poses against a white
background. No text or other graphics-just rows of flat illustration
stretching off to the right.
Without this change, the Figma, Playwright, Chrome, and other visual MCP
servers are pretty much entirely useless.
I tested this change with the openai respones api as well as a third
party completions api
This PR adds support for streamable HTTP MCP servers when the
`experimental_use_rmcp_client` is enabled.
To set one up, simply add a new mcp server config with the url:
```
[mcp_servers.figma]
url = "http://127.0.0.1:3845/mcp"
```
It also supports an optional `bearer_token` which will be provided in an
authorization header. The full oauth flow is not supported yet.
The config parsing will throw if it detects that the user mixed and
matched config fields (like command + bearer token or url + env).
The best way to review it is to review `core/src` and then
`rmcp-client/src/rmcp_client.rs` first. The rest is tests and
propagating the `Transport` struct around the codebase.
Example with the Figma MCP:
<img width="5084" height="1614" alt="CleanShot 2025-09-26 at 13 35 40"
src="https://github.com/user-attachments/assets/eaf2771e-df3e-4300-816b-184d7dec5a28"
/>
Curtis 'Fjord' HawthorneandGitHub
aaronl-openaiandGitHub
josiah-openaiandGitHub
Vivian FangandGitHub
pakrym-oaiandGitHub
Matthew ZengandGitHub
Michael BolinandGitHub
Gabriel PealandGitHub