AgentAPI modules conflict when multiple AI coding tools are installed (hardcoded port 3284)

[qinkangdeid]

Problem

All AgentAPI-based modules (claude-code, codex, auggie, gemini, opencode) hardcode agentapi_port = 3284 in their internal agentapi submodule. When multiple modules are installed in the same workspace template, only the first one to start actually works — the rest silently connect to the wrong AgentAPI instance.

For example, with both claude-code and codex enabled, clicking "Codex" in the Coder UI opens Claude Code instead.

Affected modules

• coder/claude-code (agentapi port 3284)
• coder-labs/codex (agentapi port 3284)
• coder-labs/auggie (agentapi port 3284)
• coder-labs/gemini (agentapi port 3284)
• coder-labs/opencode (agentapi port 3284)

Expected behavior

Each module should be able to run on a different port so they can coexist in the same workspace.

Proposed fix

Expose agentapi_port as a variable in each module's main.tf, then pass it through to the agentapi submodule. This is a one-line change per module:

variable "agentapi_port" {
  type        = number
  description = "The port used by AgentAPI."
  default     = 3284
}

Then in the module "agentapi" block:

agentapi_port = var.agentapi_port

Users could then configure non-conflicting ports:

module "claude-code" {
  # ...uses default 3284
}

module "codex" {
  agentapi_port = 3285
  # ...
}

module "gemini" {
  agentapi_port = 3286
  # ...
}

Environment

• Coder v2.31.7
• Template with PVE LXC backend
• All modules at latest versions

🤖 Generated with Claude Code

[matifali]

Hi there!
We are in the process of modularizing agent modules. not include agentapi within to allow users better control over how they want to build multiple agent templates.

@35C4n0r can share details on how it will work.

[qinkangdeid]

Hi @35C4n0r — following up on @matifali's note about modularizing the agent modules to decouple AgentAPI. Could you share any details on the plan?

Specifically:

1. Rough timeline — is this landing in the next few weeks, or longer?
2. Migration path — will coder-labs/codex (etc.) keep the embedded AgentAPI as a backwards-compatible default, or will users need to compose coder/agentapi + coder-labs/codex explicitly?
3. Interim fix viability — would you be open to merging fix: expose agentapi_port in codex/auggie/gemini/opencode modules #838 as a short-term patch (exposes agentapi_port so multiple AI tools can coexist today), or would you prefer we wait for the new architecture?

Context: we currently run Claude Code + Codex in the same workspace and hit the hardcoded-3284 conflict daily. Happy to adapt whichever direction helps most.

[matifali]

we will do major version bump for all affected modules os if you stay on the current version you will be unaffected.

The timeline is next few weeks. It is actively under development.

The main migration path would be the README page on module pages.

We will also add a guide or two cover use cases

1. Run multiple agents in single template
2. Enable Agent firwalls
3. Run Coder Tasks
   etc.

I am not sure about the flow of the guides but we will cover the main use cases.

If you have any other use case please share.

[matifali]

We have migrated both Claude-code and Codex modules to remove the agentAPI dependency and hence Coder Tasks integration in the refactored v5 of the modules.

Furthermore, Coder Tasks will be deprecated starting June 2nd.