Positron Assistant now has preview support for Custom Providers (OpenAI compatible)

[jthomasmock]

As of Positron 2025.10.0, we have added initial preview support for Custom Providers for Positron Assistant! This will also soon be available in preview for Posit Workbench: Positron Pro in the next Q4-2025 release.

Custom Providers allow you to use arbitrary AI model providers that adhere to the OpenAI API specification. Custom Providers could include tools such as litellm, vllm, OpenRouter, Portkey.ai, or other arbitrary centralized routing, proxying, or LLM hosting frameworks that otherwise sit in front of direct model provider access.

Note

In the future, we intend to provide additional guardrail checks for models that have the core capabilities necessary to provide a positive Positron Assistant experience, which at a minimum, would be tool-calling support. In future iterations of Custom Provider support, we intend to at least warn users if models do not meet sufficient requirements to provide the expected user experience.

Caution

While Databot can also use models via a Custom Provider, we currently only suggest using Anthropic models with Databot per our testing during its Research Preview.

Why Custom Providers?

Data Science teams

Custom Providers may be interesting for teams that require:

• On-premise or private cloud deployment: All code and prompts remain within your private or even offline infrastructure, meeting regulatory requirements (HIPAA, GDPR) without transmitting data to external services or countries.
• Governance or auditability: Additional governance tools on top of model routing to provide complete audit trails and centralized access controls. This could enable compliance demonstration and usage monitoring across the organization
• Flexibility and/or Vendor independence: Switching between model providers and model versions without a change in user workflows
• Cost and performance optimization: Deploy models to multiple users near development environments for reduced latency, predictable costs, and custom resource allocation based on organizational needs.

Individual Data Scientists

Individuals may be interested in using Custom Providers to access “free” foundational models that run in-session via tools like Ollama, which often means sharing compute with their data science environment. Our current recommendations, based on model quality and Positron Assistant’s need for tool-calling and agentic workflows, are that models should be run in a separate computational environment.

This recommendation is based on:

• Sufficient model quality: Larger models (70B+ parameters) typically deliver better code completion, reasoning, and problem-solving results compared to smaller models, at the cost of much larger RAM/GPU requirements.
• Required model capabilities: Models must support tool-calling and agentic workflows to enable the full functionality of Positron Assistant
• Resource isolation: Running models separately prevents resource conflicts with data science workloads and allows independent scaling of AI tools and analytical computation.

Enable Custom Providers

To make use of Custom Providers, Positron Assistant must first be enabled as outlined in the Positron Assistant Getting Started Guide. Custom Providers can then be enabled via the positron.assistant.enabledProviders setting in your settings.json.

While Custom Providers could be used to access OpenAI endpoints, you should instead make use of the “openai” option within the same setting. The direct OpenAI option will change to “openai-api” in the 2025.11.0 release to avoid naming conflicts we have recently observed with other tools and model providers.

"positron.assistant.enabledProviders": [
  "openai-compatible",  //## <-- Custom Providers that adhere to OpenAI API standards
  //"openai",    //## <-- For using OpenAI directly in 2025.10.0
  //"openai-api" //## <-- Planned renaming for OpenAI in 2025.11.0
]

Once the Custom Provider option is enabled, you can sign in to your custom provider.

Sign-in to Custom Provider

1. Run the Positron Assistant: Configure Language Model Providers command via the Command Palette to open the language model provider dialog. You can also click the Add Model Provider… button in the top left of Positron Assistant’s chat area.
2. Select “Custom Provider”
3. Authenticate with the model provider by supplying a:
   1. Base URL: API endpoint specific to this provider, router or proxy.
   2. API Key: A long-lived authentication token that provides access to the model service.
4. Once authenticated successfully, the Sign in button will change to Sign out, and available models from the provider will be registered. Some custom providers such as OpenRouter may take additional time to authenticate.
5. Click the Done button once you have added all the desired model providers. You can always come back to this dialog later to add or remove providers.

We want your feedback!

We've heard the interest for Custom Providers in #8592 and #8326 and are happy to provide this initial support.

Please give it a try, and leave any ideas or feedback in this discussion or open an issue if you run into any problems!

[Upipa]

Hi i have problems configuring models provided by azure foundry. I configured liked said above

"positron.assistant.enabledProviders": ["openai-compatible"]

In the configuration panel i setted the API key and the base URL (the same that i use with ellmer) but it fails with the error

Failed to add language model Custom Provider: Could not fetch models Resource Not Found

Am I doing something wrong or there's simply no support for azure foundry right now?

[sharon-wang]

Hey @Upipa, thanks for giving this a try! I think you're running into #9971 (comment) -- we try to fetch models from the /models endpoint, but from the error message you provided, it seems like Azure Foundry may not support that endpoint. Once we have a solution for that issue, it should unblock this particular error you're seeing!

These other issues may be of interest to follow, although they are focused on Azure OpenAI, which may different implementation requirements than Azure Foundry.

• Spike: Test Azure OpenAI via OpenAI different base URL + API key #10171
• Positron Assistant: Support Azure Foundry with SSO/OAuth via EntraID #10172

[Upipa]

Very sad, i was so excited to try the assistant. I'll wait for this problem to be fixed. Thanks for the answer.

[viv-analytics]

Could Databricks‘ foundational models be integrated in this way too? I'm curious about the possibilities!

[JuanVeraF]

Positron Version: 2025.11.0-234
Ubuntu 24.04

Hi. I'm having problems with OpenRouter. I was able to add this custom provider

My settings.json:

{
"files.associations": {
"renv.lock": "json"
},
"editor.hover.enabled": false,
"workbench.colorTheme": "Default Positron Dark",
"editor.parameterHints.enabled": false,
"terminal.integrated.stickyScroll.enabled": false,
"chat.agent.todoList": {
"position": "default"
},
"positron.assistant.enable": true,
"positron.assistant.enabledProviders": [
"openai-compatible",
"openai-api"
],
"positron.assistant.showTokenUsage.enable": true
}

But I don't have the option to switch to this.

Any idea what I'm doing wrong?

Thank you in advance,

Juan.

[sharon-wang]

Hi @JuanVeraF, your setup looks right 👍 I believe you're running into #10321, and we are working on a fix for this. I've copied your comment over there for reference. Stay tuned for a fix!

[mkrasmus]

Doesn't appear to work for me. I'm on version 11 with the following JSON settings:
"positron.assistant.enable": true,
"positron.assistant.enabledProviders": [
"openai-compatible", //## <-- Custom Providers that adhere to OpenAI API standards
"openai-api" //## <-- Planned renaming for OpenAI in 2025.11.0
]

"Failed to add language model OpenAI: "Model incompatible request argument supplied: temperature"

Any ideas?

[sharon-wang]

Hey @mkrasmus, you've got the correct settings 👍 I think you're running into #10469. We are actively working on a fix -- please follow along in #10469 for updates!

[henningsway]

I've been trying to use models provided by Langdock API (Anthropic), but was not yet successful with the setup.

[sharon-wang]

Hey @henningsway!

Note that the Custom Provider is for OpenAI-compatible endpoints, so this Anthropic endpoint is not expected to work. If you're up for giving this a try anyways, one additional setup step is to manually configure a model listing using the positron.assistant.models.custom setting, since Langdock does not appear to implement the /models endpoint, which Positron Assistant uses to list available models.

This may be of interest to follow for Custom Provider support for Anthropic endpoints: #8007

[henningsway]

Thank you @sharon-wang!

I've experimented with adding the following to my settings.json, but so far I've not been successful. Maybe reaching out to Langdock for information regarding OpenAI-compatability seems like a good direction

{
  "positron.assistant.models.custom": {
    "openai-compatible": [
      {
        "name": "Claude Sonnet 4.5",
        "identifier": "claude-sonnet-4-5-20250929",
        "maxInputTokens": 200000,
        "maxOutputTokens": 64000
      },
      {
        "name": "Claude Sonnet 4",
        "identifier": "claude-sonnet-4-20250514",
        "maxInputTokens": 200000,
        "maxOutputTokens": 64000
      },
      {
        "name": "Claude 3.7 Sonnet",
        "identifier": "claude-3-7-sonnet-20250219",
        "maxInputTokens": 200000,
        "maxOutputTokens": 8192
      },
      {
        "name": "Claude 3.5 Sonnet",
        "identifier": "claude-3-5-sonnet-20240620",
        "maxInputTokens": 200000,
        "maxOutputTokens": 4096
      }
    ]
  }
}

[sharon-wang]

@henningsway do you have access to OpenAI models with Langdock?

If so, I think you should be able to use the base url https://api.langdock.com/openai/eu/v1, along with configuring positron.assistant.models.custom with the supported models from https://docs.langdock.com/api-endpoints/completion/openai -- perhaps: gpt-5 and gpt-4.1 ?

There's this note as well:

The list of available models might differ if you are using your own API keys in Langdock (“Bring-your-own-keys / BYOK”, see here for details). In this case, please reach out to your admin to understand which models are available in the API.

[henningsway]

Well, actually this works. :) Thank you for getting back to me.

[jarbet]

Is there a plan to support Claude's Pro plan ($20/month) rather than requiring an API key which has a separate pricing system (I think)? Claude's Pro plan now supports Claude Code too.

It be nice to have an option to use something like Claude and not have to worry about getting a large bill but rather one would pay a fixed amount per month.

[sharon-wang]

Hey @jarbet, we are tracking this request in #9481. If you'd like, upvote (👍) and follow along there for updates!

We hear you and other folks on the pricing/billing note for Anthropic -- the current limitation is that Claude Pro does not support API access, which tools like Positron Assistant need to provide the chat experience. Although the Claude Pro plan gives access to Claude Code, it does not give access to API access to other AI agents like Positron Assistant. If Anthropic allows API access through Claude Pro or another plan in the future, that would alleviate this main blocker!

[ChangenetAlex]

Hi all, thanks a lot for the work.
Just wanted to say this works nicely for me on Positron Version: 2026.02.1 build 5. I can now use the deployed models from our azure foundry instance.

If this can be helpful, it took me a while to figure out but to specify a custom model:

• The config option in settings.json is not "positron.assistant.configuredModels": {} as I read somewhere else.
• But the config option is "positron.assistant.models.custom": {} as read in this discussion.

[juliasilge]

In the 2026.03 release, some of these changed FYI. Check out our updates docs:

• https://positron.posit.co/assistant-chat.html#managing-models
• https://positron.posit.co/assistant-getting-started.html#step-2-configure-language-model-providers under "Custom Providers"

[ChangenetAlex]

Thanks. I updated to latest version and settings.jsonconfigs were updated automatically. All work fine