What version of Codex is running?
codex-cli 0.125.0
Platform
macOS 26.4.1, Build 25E253
Problem
Codex desktop/app-server appears to repeatedly spawn duplicate STDIO MCP server processes for configured MCP servers and does not clean up old instances.
In this case, two configured STDIO MCP servers both use:
command = "uv"
args = ["run", "mcp-atlassian"]
Over time, process count grows substantially. During investigation, the count grew from 48 processes (~1.9 GB RSS) to 60 processes (~2.27 GB RSS), without explicit use of the Jira MCP tools.
Observed process pattern
Each server instance appears as a pair:
uv run mcp-atlassian- Python child running
~/.local/bin/mcp-atlassian
All are parented under Codex app-server:
Codex.app/Contents/Resources/codex app-server --analytics-default-enabled
Example process groups were started at different times, indicating repeated spawning rather than a single long-lived pair:
- 4:43 PM
- 5:22 PM
- 7:21 PM
- 7:23 PM
- 7:35 PM
- 7:39 PM
Expected behavior
Codex should either:
- keep one live STDIO MCP server process per configured MCP server per active Codex client/session, or
- clean up previous MCP server children when a session/tool client is closed/reloaded, or
- provide an idle timeout / lazy-start mode for MCP servers.
Actual behavior
Duplicate MCP server processes accumulate and remain running. They consume memory and small but persistent CPU even when tools are not actively being called.
Relevant config shape
Sanitized:
[mcp_servers.jira-dev]
command = "uv"
args = ["run", "mcp-atlassian"]
startup_timeout_sec = 120.0
[mcp_servers.jira-sd]
command = "uv"
args = ["run", "mcp-atlassian"]
startup_timeout_sec = 120.0
Notes
This does not appear to be the MCP server intentionally doing background Jira work. The installed MCP server mostly registers tools and waits on STDIO. The issue seems to be repeated Codex-side spawning and/or missing cleanup.
What version of Codex is running?
codex-cli 0.125.0Platform
macOS 26.4.1, Build 25E253
Problem
Codex desktop/app-server appears to repeatedly spawn duplicate STDIO MCP server processes for configured MCP servers and does not clean up old instances.
In this case, two configured STDIO MCP servers both use:
Over time, process count grows substantially. During investigation, the count grew from 48 processes (~1.9 GB RSS) to 60 processes (~2.27 GB RSS), without explicit use of the Jira MCP tools.
Observed process pattern
Each server instance appears as a pair:
uv run mcp-atlassian~/.local/bin/mcp-atlassianAll are parented under Codex app-server:
Codex.app/Contents/Resources/codex app-server --analytics-default-enabledExample process groups were started at different times, indicating repeated spawning rather than a single long-lived pair:
Expected behavior
Codex should either:
Actual behavior
Duplicate MCP server processes accumulate and remain running. They consume memory and small but persistent CPU even when tools are not actively being called.
Relevant config shape
Sanitized:
Notes
This does not appear to be the MCP server intentionally doing background Jira work. The installed MCP server mostly registers tools and waits on STDIO. The issue seems to be repeated Codex-side spawning and/or missing cleanup.