Codex and opencode on Uno AI

Read this first. Codex is built around OpenAI's models; opencode is built to work with any model. Both work on Smart and Fast through Uno, and tool calls can fail where they would not on the model each was built for. On 16 October 2026 we pointed Claude Code 2.1.284, Codex 0.154.0 and opencode 1.18.32 at Uno with a test key: each reached Uno's endpoint and got Uno's own answer to a wrong key. On 25 October 2026 opencode inside Uno Work ran a full task on Uno AI with tool calls: it wrote, started and tested a small API. A run of the local CLIs with a real key is still to do.

Steps

1. Copy your Uno key and keep it in a variable

   Open the Uno console: Uno AI, then Use Uno AI from your code and copy the API key. Both agents read it from UNO_API_KEY, so the key never sits in a config file. Add the line to ~/.bashrc to keep it.

   $ export UNO_API_KEY=your-uno-key

2. Test the key with one request

   An answer with "choices" means the address and the key work.

   $ curl -sS https://console.uno.place/v1/chat/completions \
       -H "Authorization: Bearer $UNO_API_KEY" \
       -H "Content-Type: application/json" \
       -d '{"model": "uno/fast",
            "messages": [{"role": "user", "content": "Say hi"}]}'

3. Codex: add Uno as a provider

   Add this block to ~/.codex/config.toml. It only defines the provider: your ChatGPT sign-in stays the default. Recent Codex versions take only the Responses format from a custom provider, so keep wire_api = "responses".

   [model_providers.uno]
   name = "Uno AI"
   base_url = "https://console.uno.place/v1"
   env_key = "UNO_API_KEY"
   wire_api = "responses"

4. Codex: run it on Uno AI

   Choose the provider and the model when you start Codex. The same flags work with codex exec for a task without the interface.

   $ cd your-project
   $ codex -c model_provider='"uno"' -m uno/smart
   # one task, no interface:
   $ codex exec -c model_provider='"uno"' -m uno/fast "Fix the failing test"

   To make Uno the default, put model_provider = "uno" and model = "uno/smart" at the top of the file. Codex warns that it has no metadata for the model and uses defaults: the run goes on.

5. opencode: add Uno as a provider

   Put this in opencode.json in the project folder, or in ~/.config/opencode/opencode.json for all projects. opencode takes the key from the same UNO_API_KEY.

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "uno": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Uno AI",
         "options": {
           "baseURL": "https://console.uno.place/v1",
           "apiKey": "{env:UNO_API_KEY}"
         },
         "models": {
           "uno/smart": { "name": "Smart" },
           "uno/fast": { "name": "Fast" }
         }
       }
     }
   }

6. opencode: pick the model

   Start opencode and type /models. Choose Smart or Fast under Uno AI. For one task from the command line, name the model as provider, then model:

   $ opencode
   $ opencode run -m uno/uno/fast "Rename getUser to fetchUser everywhere"

Check that it works

Give each agent a small task that edits one file, then open the Uno console. The minutes appear on the Uno AI page. only minutes the AI works count; ten chats in the same minute count as one, so Codex and opencode working side by side spend one set of minutes.

$ codex exec -c model_provider='"uno"' -m uno/fast -s workspace-write "Add a line to README.md: checked on Uno AI"
$ git diff README.md

Codex or opencode on Uno AI

What it does not do

• It does not turn Smart and Fast into GPT. On the Artificial Analysis index Smart has 46, GPT-6.1 Sol has 52.
• The Responses endpoint keeps no state between calls: Codex sends the whole conversation each time, which is how it works with other providers too.
• A new Codex version can change what it sends and break the setup until Uno catches up.
• On Small+AI and Plus+AI the work stops when the hours end. Only Max+AI is unlimited.

Common problems