Pool stdio MCP connections between tool calls

[RhysSullivan]

Every tool call on a stdio MCP integration spawned a fresh child process, ran the full MCP handshake, called one tool, and tore the child down. Measured with @modelcontextprotocol/server-everything on the local daemon: 875–897ms on every call of a no-op echo, flat. This is the desktop/local host only — cloud disables stdio — and it is the dominant per-call overhead for anyone running npx-launched servers.

Plain stdio now joins the existing connection pool that remote and app-server connections already use: same five-minute idle window, same hashed identity key, same drop-on-transport-failure semantics. Identity separates on command, args, cwd, secret env, credential values, template, owner, and connection, and plain stdio never shares a session with the app-server bridge. Per-call elicitation handlers were already rebound on each pooled use, so reuse changes no elicitation behavior.

Opt-out: spawnPerCall: true in the stdio config (also accepted by the add-server API) restores fresh-process-per-call for servers that depend on it. The app-server bridge ignores the flag — its approvals are session state.

Measured on this branch, same server and tool:

	before	after
stdio echo, first call	~880ms	585ms (spawn + handshake, then parked)
stdio echo, subsequent calls	875–897ms each	3.5–8ms
pooled children during a burst	n/a	exactly 1
children after server shutdown	—	0 (pool closes with the plugin)

New tests pin the poolability predicate (default on, opt-out honored, bridge always pooled) and the stdio key separation properties (secret env, command/args/cwd, bridge-vs-plain, owner).

Codex plugin compatibility (checked against #1837/#1858/#1859): curated plugins ride the app-server bridge and were pooled before this change — no behavior change. Scanned cache plugins spawn their own MCP server binary and now pool like any stdio server; their tests pass and none rely on fresh-process semantics. One noted shift: after a plugin update moves the versioned cache dir, a parked child keeps serving until the idle window or a transport failure drops it, so the "command missing" health signal appears on the next fresh dial rather than on the very next call.

[github-actions]

Cloudflare preview

Torn down — the PR is closed.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	executor-marketing	5632b28	Commit Preview URL Branch Preview URL	Aug 30 2026, 12:11 AM

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
✅ Deployment successful! View logs	executor-cloud	5632b28	Aug 30 2026, 12:12 AM

[pkg-pr-new]

Open in StackBlitz

@executor-js/cli

npm i https://pkg.pr.new/@executor-js/cli@1861

@executor-js/config

npm i https://pkg.pr.new/@executor-js/config@1861

@executor-js/execution

npm i https://pkg.pr.new/@executor-js/execution@1861

@executor-js/sdk

npm i https://pkg.pr.new/@executor-js/sdk@1861

@executor-js/codemode-core

npm i https://pkg.pr.new/@executor-js/codemode-core@1861

@executor-js/runtime-quickjs

npm i https://pkg.pr.new/@executor-js/runtime-quickjs@1861

@executor-js/plugin-file-secrets

npm i https://pkg.pr.new/@executor-js/plugin-file-secrets@1861

@executor-js/plugin-graphql

npm i https://pkg.pr.new/@executor-js/plugin-graphql@1861

@executor-js/plugin-keychain

npm i https://pkg.pr.new/@executor-js/plugin-keychain@1861

@executor-js/plugin-mcp

npm i https://pkg.pr.new/@executor-js/plugin-mcp@1861

@executor-js/plugin-onepassword

npm i https://pkg.pr.new/@executor-js/plugin-onepassword@1861

@executor-js/plugin-openapi

npm i https://pkg.pr.new/@executor-js/plugin-openapi@1861

executor

npm i https://pkg.pr.new/executor@1861

commit: 5632b28