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.

[henningsway]

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

[fh-mthomson]

Thank you! positron.assistant.models.custom was also required to query from models hosted via Databricks Serving Endpoints:

Base URL: https://dbc-myinstancexyz.cloud.databricks.com/serving-endpoints

{
    "positron.assistant.enable": true,
    "positron.assistant.enabledProviders": [
        "openai-compatible"
    ]
    "positron.assistant.models.custom": {
        "openai-compatible": [
        {
            // model on AWS Bedrock served via Databricks
            "name": "Claude Sonnet 4.6",
            "identifier": "claude-sonnet-4-6",
        },
        {
            // model served natively via Databricks
            "name": "GPT 5",
            "identifier": "databricks-gpt-5"
        }
    ]
  }
}

[JT-39]

I can't seem to get this to work for databricks! :/ any suggestions?

[juliasilge]

@JT-39 It will be helpful to start a separate, new discussion post with all your details (Positron version, OS, etc) so we can help you get to the bottom of this!

[JT-39]

@juliasilge I've created a new discussion here (hopefully this is what you meant 😃

#12798

[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

[PanSX-Dr]

{
"files.associations": {
"renv.lock": "json"
},
"positron.r.customRootFolders": [
"C:/Program Files/R"
],
"positron.r.customBinaries": [
"C:/Program Files/R/R-4.4.3/bin/x64/R.exe"
],
"workbench.keybindings.rstudioKeybindings": true,
"json.schemas": [],
"positron.r.pipe": "%>%",
"workbench.colorTheme": "Default Positron Dark",
"positron.assistant.enable": true,
"positron.assistant.enabledProviders": ["openai-compatible"],
"positron.assistant.models.include": [
"^gpt-4o-mini$",
"^gpt-4o$",
"^gpt-5\.2$"
],
"positron.assistant.models.preference.global": "gpt-4o-mini",
"editor.fontSize": 16,
"positron.assistant.provider.customProvider.enable": true
}
Hello, I've been setting it up in Positron for a long time, but it keeps prompting me that it can't add a language model. What could be the problem?

[afuetterer]

Hi, I tested this successfully with two openai compatible custom providers, I have access to.

Two related questions:

• Is there a way to rename the custom provider?
• Is there a way to add multiple custom providers to switch between (right now I need to log out of provider 1, then log in to provider 2 etc) <- this is also why renaming the providers would be nice

Any help is appreciated.

[melissa-barca]

Hi @PanSX-Dr thank you for trying this out. The error message indicates that Assistant was able to authenticate with the provided credentials, but it couldn't find any models for the provider. I believe your positron.assistant.models.include regexes are too restrictive. Can you try modifying them to remove the characters ^ and $ to allow for any prefixes or suffixes that may be appended? If this doesn't work generalizing even further could help as a test, such as to gpt-4o or gpt.

[melissa-barca]

Hi @afuetterer I'm glad to hear this is working for you! Unfortunately, neither renaming custom providers nor configuring multiple custom providers are currently supported. These are reasonable requests, would you mind filing issues for them?

[afuetterer]

Will do.

[lindahu99]

Hi @jthomasmock , I'm currently using Positron Pro on Posit Workbench. I was wondering if there’s any way to configure a custom AI provider for positron assistant, given that you mentioned it would be released in Q4 2025.

[juliasilge]

This depends on exactly which version of Workbench and Positron you have installed on your server. Can you take a look at these docs:
https://docs.posit.co/ide/server-pro/admin/positron_sessions/positron-assistant.html

Check out the "Custom models" section there, but be sure to use the version selector to match which versions of Workbench and Positron you are running.

[lindahu99]

I am using Posit Workbench 2026.01.2+418.pro1 and Positron Version: 2026.02.1 build 5. I'm still getting error when using this custom model configuration.

[melissa-barca]

You'll also need to enable openai-compatible. I believe for your version of Positron this looks like the following, though this setting name changes in a later version of Positron:

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

[lindahu99]

Thanks! It works now.

[uhkeller]

I'm trying to get DeepSeek's OpenAI-compatible API to work, but with not luck. Any help is greatly appreciated. I'm using Positron 2026.04.1 build 10.

Adding

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

to the settings results in a message saying this setting is deprecated. I have

"positron.assistant.provider.customProvider.enable": true

and can add DeepSeek as a custom provider. I can also select the model (deepseek-v4-pro) in the chat dialog. But when I try to actually do something I get this message in the chat pane:

[deepseek-v4-pro] Error in chat response:
{
  "error": {
    "message": "The reasoning_content in the thinking mode must be passed back to the API.",
    "type": "invalid_request_error",
    "param": null,
    "code": "invalid_request_error"
  }
}

Here's the diagnostic output:

## Chat Customization Diagnostics
*WARNING: This file may contain sensitive information.*

**Custom Agents**<br>
*1 file loaded*

.github/agents<br>
❌ .vscode/positron/agents - *Folder does not exist*<br>
User Data<br>
Extension: GitHub.copilot-chat<br>
└─ [`Plan.agent.md`](../../../Library/Application%20Support/Positron/User/globalStorage/github.copilot-chat/plan-agent/Plan.agent.md)<br>

**Instructions**<br>

.github/instructions<br>
❌ .vscode/positron/instructions - *Folder does not exist*<br>
User Data<br>

**Prompt Files**<br>
*3 files loaded*

.github/prompts<br>
❌ .vscode/positron/prompts - *Folder does not exist*<br>
User Data<br>
Extension: GitHub.copilot-chat<br>
├─ [`savePrompt.prompt.md`](../../../../../Applications/Positron.app/Contents/Resources/app/extensions/positron-copilot-chat/assets/prompts/savePrompt.prompt.md)<br>
├─ [`plan.prompt.md`](../../../../../Applications/Positron.app/Contents/Resources/app/extensions/positron-copilot-chat/assets/prompts/plan.prompt.md)<br>
└─ [`init.prompt.md`](../../../../../Applications/Positron.app/Contents/Resources/app/extensions/positron-copilot-chat/assets/prompts/init.prompt.md)<br>

**Skills**<br>

.github/skills<br>
.agents/skills<br>
.claude/skills<br>
~/.copilot/skills<br>
~/.agents/skills<br>
~/.claude/skills<br>

[sharon-wang]

Hey @uhkeller, I think you're running into #13211 -- please follow along there for updates! It seems like DeepSeek models were previously working with Custom Provider, but there was a recent change to DeepSeek models that now requires reasoning_content to be included in responses.

[uhkeller]

Ah, sorry I missed this @sharon-wang, and thanks for pointing it out!