Configure the server#
Use startup options to select the target conda installation and bound each child process. These settings remain fixed until the MCP server restarts.
Select conda explicitly#
Pass an executable path when the MCP server and target conda live in different environments:
ccm --conda-exe /absolute/path/to/conda
Resolution follows this order:
the value passed to
--conda-exethe
CONDA_EXEenvironment variablecondafound onPATH
GUI-launched clients often receive a smaller PATH than an interactive shell.
Prefer an absolute path in client configuration. On Windows, point to
conda.exe instead of a .bat or .cmd launcher.
Set time and data limits#
The default timeout is 900 seconds. The server retains at most 1 MiB from each output stream and accepts at most 1 MiB of raw-tool input.
Set different positive limits at startup:
ccm \
--conda-exe /absolute/path/to/conda \
--timeout 120 \
--output-limit-bytes 2097152 \
--stdin-limit-bytes 65536
--output-limit-bytes applies separately to standard output and standard
error. The child process is still drained after the retained limit is reached.
The result marks each truncated stream independently.
The raw conda_execute tool may shorten the timeout for one call, but cannot
raise the server limit. Generated tools use the server limit. Oversized input
and invalid working directories are rejected before conda starts.
Account for process limitations#
The server passes argv directly to conda without a shell. Shell interpolation, pipes, redirection, aliases, and shell functions are unavailable.
The server does not allocate an interactive terminal. Use conda’s noninteractive flags for commands that could prompt. Only the raw tool accepts bounded standard input and a working directory.
Activation commands can return shell code or environment changes, but cannot modify the MCP host application’s shell. The client must apply returned changes in its own shell process.
See the server CLI reference for every option.