Claude deployments on Microsoft Foundry aren't reachable from Assistant

[codeswitching]

I know MS Foundry support in Positron Assistant is still brand new, so maybe the team is already aware of/working on this issue. But I want to highlight it because my team uses Claude through Foundry rather than direct through Anthropic (we're government).

System details

Positron 2026.08.1 b2
Windows 11

Description

I have a Claude model deployed in a Foundry resource, alongside model-router. I can't get Assistant to talk to it directly.

This seems to be a combination of two subissues:

1. Positron Assistant is hard-wired to call only model-router through Foundry; all other deployments are inaccessible. Even if model discover isn't supported in Foundry, there should be some way to specify a different model: in providers.json.*

2. Claude deployments on Foundry only speak the Anthropic Messages api (/anthropic/v1/messages), while Positron Assistant can only make OpenAI-compatible calls when talking to Foundry. When Assistant is pointed at an Anthropic-style endpoint, it authenticates, but fails at the first query with "Foundry endpoint not found. Check your base URL."

I also tried pointing Assistant's separate Anthropic provider at services.ai.azure.com/anthropic/v1/messages instead, since it should already speak the right protocol... Sign-in fails silently, no error. Maybe some difference in how the api key is passed?

Steps to reproduce

1. Deploy a Claude model in a Foundry resource.
2. Configure Language Model Providers -> Microsoft Foundry -> sign in with the /openai/v1 base URL. Succeeds, but only model-router shows up.
3. Try signing in to Foundry with the Anthropic-style base URL instead -> query -> "Foundry endpoint not found."
4. Delete the model-router deployment, cleverly redeploy your Claude model as model-router, sign back in, send a query → 404 Requested API is currently not supported.
5. Switch to Anthropic provider, for base URL use https://<resource>.services.ai.azure.com/anthropic/v1, then sign in via the Anthropic provider with an Azure API key → fails silently.

Expected

Some way to reach a specific Claude deployment on Foundry directly — either the Foundry provider supporting the Anthropic Messages protocol, or the Anthropic provider detecting an Azure host and authenticating correctly.

• It's true that the Foundry model-router can technically route to a Claude deployment, but the current model-router version (2025-11) doesn't support Azure-hosted v2 Claude deployments and lags behind current Claude releases generally — so it can't get you to newer models even though the connection itself is fine.

[melissa-barca]

Hey @codeswitching you can specify custom models in providers.json, and that is how we expect users to configure alternatives to model-router for Foundry. Foundry doesn't support querying for available models so they need to be predefined somewhere.
That being said, I think you want to define this as Anthropic provider as the Foundry provider only makes OpenAI-compatible calls. For models, you'll need to set "discovery": "off" to prevent Positron from attempting to query for available models and then define the models you want to use. Your providers.json should look something like this:

  {
    "version": 1,
    "providers": {
      "anthropic": {
        "baseUrl": "https://<resource>.services.ai.azure.com/anthropic/v1",
        "models": {
          "discovery": "off",
          "custom": [
            {
              "id": "my-claude-deployment",
              "name": "Claude (Foundry)",
              "maxContextLength": 200000,
              "supportsTools": true,
              "supportsImages": true,
              "supportsToolResultImages": false,
              "supportsWebSearch": false
            }
          ]
        }
      }
    }
  }
  }