Option to disable ALL built-in tools? (please?)

[fasterthanlime]

Hello Zed friends!

For reasons I would like an option to disable every single built-in ACP tool available to Claude.

I provide my own, and they keep trying the built-in ones, running afoul of sandboxing limitations etc. — denying permission is not enough because they just, won't, stop, trying. Stubborn little things.

So, yeah, a --disable-all-builtin-mcp-tools would be amazing — hopefully it would also amend the agent's prompt so it doesn't even learn about these built-in tools in the first place.

I'm running over to claude-code-acp to open a similar issue, so, if this reads familiar, that's why.

[benbrandt]

@fasterthanlime codex-related topics here that I didn't cover in the claude code one:

• This is interesting for codex because we don't really provide it tools... and other than MCP servers, codex doesn't model a lot of this as "tools" per-se the same way that claude code does.
• Most tool calls for codex are actually just shell commands, and then they parse the commands to make them more readable
• Codex also already does a bunch of sandboxing... is this not enough? Are you doing your own sandboxing that doesn't play well with theirs?
• Is the issue that you have MCP servers configured/installed in Zed that you don't want to send?

[fasterthanlime]

• This is interesting for codex because we don't really provide it tools... and other than MCP servers, codex doesn't model a lot of this as "tools" per-se the same way that claude code does.

Ah! That would make sense. codex is super fond of "cat a HEREDOC into a file" (the polar opposite of gemini, which uses the ReadManyFiles tool lol).

Any way to disable "shell commands" altogether? If not here perhaps at the codex CLI level directly? I understand I could give it a sandbox profile where every exec essentially fails but it's not gonna stop it from trying.

• Is the issue that you have MCP servers configured/installed in Zed that you don't want to send?

No no, just that I want every single thing it does (besides reasoning) to go through my MCP server.

[benbrandt]

Any way to disable "shell commands" altogether?

Hmm I'm not seeing anything obvious from the config. Because literally everything is going through the terminal, it is pretty baked into the agent loop.

The only hooks I see are some stuff for adding some initial conversation history items in your AGENTS.md file, or overwriting the base instructions.

Here's the system prompt (there are also some model specific additions here and here that you can look at.

I think you can use this flag to provide a different one? https://github.com/zed-industries/codex/blob/fa4cd7ed15b280f2a2e250531d0a7cdb1f50db64/codex-rs/core/src/config/mod.rs#L654

Or somehow get this config item to change: https://github.com/zed-industries/codex/blob/fa4cd7ed15b280f2a2e250531d0a7cdb1f50db64/codex-rs/core/src/config/mod.rs#L129