Introducing Positron Assistant!

[jthomasmock]

Note

The Positron Assistant is currently an experimental feature in Positron, available in 2025.06.0-167 and beyond. We are actively improving the user experience and functionality, so please provide feedback and report any issues you encounter.

Positron Assistant currently supports:

• Anthropic Claude for Chat-only
• GitHub Copilot for code completions only

We are planning to expose additional model provider backends over time.

Refer to the Positron Assistant User Guide for detailed information on how to set up and use Positron Assistant.

Positron Assistant is an AI client that provides LLM integration within Positron, both for chat and for inline completions. Use Positron Assistant to generate code, ask questions, get help with debugging, refactor, and get suggestions for next steps in your data science projects.

Why use Positron Assistant over other AI clients?

The Positron Assistant offers novel data science and Positron-specific context compared to other IDE AI tooling. We provide a data-science-centric IDE for the user and allow the LLM to interact with the most important parts of Positron to get better context for better results!

A majority of data science work happens interactively with live data during development. Users load local or remote data, seek to understand it via cleaning, transformations, visualizations, and then modeling, then ultimately generate reproducible code. While the full code on disk is the ultimate source of truth and gives you the ability to reproduce the results, the ephemeral state is where a data scientist is actually working the majority of the time. As such, existing AI clients that only have access to the code on disk are limited in their ability to provide useful suggestions and context-aware completions for data science.

We use tool-calling heavily within Positron to allow the LLMs to directly interact with Positron’s novel data science features. We have built out key integration points that allow the LLM to parse and interact with the current state of the Positron IDE, and will continue to expand these capabilities over time.

A few examples of these integration points include Positron’s:

• Variables pane, providing names, dimensions, and column names of dataframes and Python/R based database connections, as well as metadata for arrays, lists, and other objects in memory
• Plots pane for describing the current plot or suggesting variations and next steps for iterating on the current plot
• Console, available to the model for viewing inputs and code outputs, as well as allowing the user or an AI agent to execute code in the console and view the output
• Session state, providing what language is active in the console session (R or Python, including version), what packages are loaded, and what packages are installed

All of that Positron-specific context is in addition to the core capabilities of what traditional software-oriented GenAI clients have (active files, selected lines of code, files in project, training data in the model, etc).

Configuring Positron Assistant

To learn more about the Positron Assistant, including turning it on and adding Model Providers, please see the Positron Assistant User Guide.

[ndiquattro]

Excited about this!

Do you know if it's in the near-term to also enable chat with Copilot? I'm trying to decide whether to go through some rigamarole to get an anthropic API key or just wait for further copilot integration. Thanks!

[jmcphers]

The Anthropic API uses token (use based) pricing. See their API pricing here: https://www.anthropic.com/pricing#api

[ntluong95]

I meant I want to know about the actual price when using, like with 1M token, what I can usually do?

[jmcphers]

It's difficult to be precise, but very roughly speaking you can think of each conversation with the assistant as consuming a few thousand tokens. I use Assistant multiple times a day (as part of developing it) and don't go through more than ~$10 in credits a month for Positron-initiated use.

[ndiquattro]

I navigated the rigamarole and acquired a key! One follow-up question, I was instructed to avoid the Files API, since that route isn't covered by our agreement for data privacy with Anthropic. Does assistant use the Files API in any contexts? Thanks again!

[juliasilge]

It does not! You can confirm for yourself by perusing this file to see how we call the Anthropic API.

[tomasrei]

Hi! Great work, Looking forward to trying this!

Currently, when using vscode, I get access to github copilot, claude (and I believe some other AI APIs) via my PRO github account. Is it possible to activate the chat and code completion via a github PRO account only, or do I need a claude account as well? Thanks.

[juliasilge]

Can you say more about what you mean here:

However it doesnt appears in the list of providers in the in-line completion

Are you meaning that you do not see this in the bottom right when you click on the robot?

[roimartins]

In the bottom right it looks Github Copilot is registered :

However in the Ctrl + I , in line interface is not listed any Language Model I can pick . When i click on "Add language model" I can see "github Copilot" options igned but when i select it and press "Done" , the inline interface still do not provide any option i can choose.

[juliasilge]

Ah, when you use Ctrl/Cmd+I in your file, that is actually a chat feature. The completions are the "ghost text" that gets generated in your file that looks like this:

You can read this VS Code documentation to get a better understanding of what the features of completions are: https://code.visualstudio.com/docs/copilot/ai-powered-suggestions

[roimartins]

ah , i got it . The chat features arent yet avaialble , is the autocompletion what works atm . Looking forward for chat features , it looks very useful extension . So far only good surprises with Positron .

[juliasilge]

The chat features are available, but not with GitHub Copilot. Right now, what we support is

• Anthropic for chat only
• GitHub Copilot for code completions only

We are planning to expose additional model provider backends as we move forward!

[davidrsch]

I tried using a local model via Ollama, but couldn’t get it to work. I’m curious, will there be official support for Ollama in the near future? Or is Posit more likely to wait for GitHub Copilot to be open sourced and instead focus on developing tools tailored for data science?

One feature I’d really love to see in Positron Assistant is the ability to be able to set it to use vignettes and package documentation as contextual references. For example, I’ve recently been working on implementing a custom engine for tidymodels, and examples on this are quite limited. In my experience, LLM assistants like ChatGPT, Gemini, or even Copilot in VS Code often fall short in these niche R scenarios, I’ve ended up completing some tasks faster manually.

Being able to ground responses in actual package docs or vignettes could make a huge difference for complex workflows like this. I think it’d be a very compelling feature.

[juliasilge]

We don't yet officially support Google Gemini and are tracking that support in #6743 and related issues.

[DDorch]

Hi @davidrsch -- as far as model providers, we do not currently support Ollama and are unlikely to prioritize local models in the short term, as models with sufficient quality generally consume too much RAM at this time.

Medium-term we are planning to add support for other providers such as Google Gemini Code Assist.

Longterm we are also planning on supporting internally hosted models through tools like vllm.

As far as the other feedback:

One feature I’d really love to see in Positron Assistant is the ability to be able to set it to use vignettes and package documentation as contextual references

Yes, we agree! There's currently an issue open for working with web-based documentation: #7649 and I've also opened: #8016 for your feature request.

Local models can be used on a dedicated huge machine for a small community that don't want to provide all its code to an external provider... So, I think that supporting local/independant LLMs as ollama would be very appreciated :)

[nviets]

Agree completely with @DDorch. In an enterprise context, air gapping networks to protect sensitive data and code is common, so cloud providers are not an option. As @DDorch mentions, huge machines are generally available within the air gap to host large high quality models. Ollama support would be greatly appreciated, @juliasilge and @jthomasmock.

[jthomasmock]

Hi @nviets and @DDorch -- thanks for the thoughts.

We have definitely prioritized internally hosted or proxied models and other arbitrary proxy/routed models via #8592 . Importantly, this will open up arbitrary URLs for your teams to send/receive chats, prompts, and responses back and forth against. The emphasis here is more of a suggestion from myself and others to separate the interactive session's compute from the LLM's in-memory overhead. I would not recommend running in-session Ollama + interactive R/Python sharing the same memory/CPU/GPU for example, although it can be done with smaller models (likely with lower quality results) or with much larger compute (that is still competing with your analysis environment resources).

The reasoning behind not directly providing support for Ollama in-session is that many models are insufficiently powered or have insufficient support for tool-calls to be capably used to power Positron Assistant and/or Databot or other future tooling we will provide. Without tool-calls, Positron Assistant is not really functional. With smaller models, many responses we have observed will be poor quality.

However, the Custom Provider endpoint we have added would allow a centrally-managed LLM server to be used that scales to multiple users and importantly scales independently of the R/Python environment where your analysis would occur. This would allow for high-quality LLMs to be served to multiple users and efficiently scale while still providing the data security model many customers wish for. In theory, Ollama can be used to provide such a service, but more likely is tools like Portkey.ai, vllm, litellm, Openrouter, etc that are intended to be used with multiple users.

[nviets]

Thanks @jthomasmock - This is perfect. Looking forward to the release.

[johnryan412]

Is there a way to add a delay before copilot attempts to inline complete your code in positron? Would be helpful if it would wait 2-3 seconds.

[testlabauto]

There is not currently a way to do this with Positron. Can you elaborate a little on why this would be useful?

[johnryan412]

When I'm typing out code, I don't always need it to guess what I'm about to type next. It becomes intrusive. RStudio has an option "Show code suggestions after keyboard idle (ms)" where you can set a delay before copilot will provide suggestions which is nice.

[testlabauto]

Oh, ok. That sounds like a good idea. Filed an enhancement for this: #8066

[PaulC91]

Similar to this, is there a way to disable automatic completions and instead invoke them with a keyboard shortcut? I can achieve this in VScode by setting "github.copilot.editor.enableAutoCompletions": false and having a keybinding for the copilot "Trigger Inline Suggestion" command. Thanks!

[juliasilge]

Not currently but we are tracking something along these lines in #8017. I added your use case there in a comment, but please feel free to add any additional info. Thanks!

[abduazizR]

Not sure if this is relevant, but can I use positron assistant to generate git commits?

[testlabauto]

It cannot currently do this, but it is a great idea. I filed an enhancement: #8072

[ntluong95]

Will the Positron Assistant be able to create project, says "Creating a shinyApp from prompt", or getting input as the user's CSV file and the prompt (could be README file, describing research questions or project aims), it could create necessary files and code?

[juliasilge]

We definitely have plans of this nature in mind! If you are specifically interested in generating Shiny apps, you might want to check out Shiny Assistant in the meantime: https://shiny.posit.co/blog/posts/shiny-assistant/

[ntluong95]

I am aware of Shiny Assistant, that's why I am asking. Having a way to integrate that into Positron is great.

[jarbet]

Any plans to support Google Gemini API for chat in the future? They offer a free tier :)

[juliasilge]

Yep, we are tracking this feature in #6743, if you want to give a 👍 or similar.

[jarbet]

Yep, we are tracking this feature in #6743, if you want to give a 👍 or similar.

Done!

In the meantime, FWIW I'll share my current Positron LLM setup in case it helps others:

• I use Positron Assistant with Github Copilot since I get this for free as an Academic. Thus I can use Copilot for code completion.
• I use the Continue Extension with Google Gemini API (for free) to ask questions about my code and help edit code.

[Ulaskarakoc]

I just played with Assistant using Claude API.
Turned out few questions cost me 200 k tokens! (0.5 dollars) I only added one short R script as context in a project folder which includes csv files, about 20 R files etc.
Is it possible that Positron Assistant sent all R files and possibly data files to Claude?

[jmcphers]

No, it doesn't send the whole project up to Claude unless you tell it to (or it queries it in Agent mode). Many of those are system prompts, internal tools, etc.; Claude in general just uses a lot of tokens.You can get some more detailed information by going to the Output panel > Assistant where it'll tell you exactly what's going up and roughly how many tokens are involved.

Also, what version of Positron are you using? In later versions we've added support for caching with Anthropic's API which should reduce token usage on subsequent questions.

[Ulaskarakoc]

I use the latest version of Positron (2025.07.0). When I add context, the token use is ridiculously high, despite the small size of R file added as context.
I asked a simple question related to R file when current file is on, the total token use is about 17-18 k. Then I turned off the current file view, then again, simple question costs 7-8 k.
All these simple 4-5 exercises cost 0.19 dollars. Is that all normal?
At this rate, I think I will keep using my normal Claude Pro and keep going back and forth between Claude app and Positron.

[sharon-wang]

@Ulaskarakoc do you recall if Assistant was making tool calls as part of its response?

[Ulaskarakoc]

I do not think so.

[ntluong95]

Testing the simple function Fix/Explain from Console today. Tried to make simple error like librari(tidyverse). It took the unnecessary context file and console session, so after 5 trials, it consumed 47k token out and 1200 token in, which cost $0.4. The issue was describe in #9209

[zachcp]

Fantastic work! Can we customize the mcp tools available to the Assistant?

[jmcphers]

Yes, you can. Assistant can use MCP tools just like Copilot can, and most of the UX is the same.

https://code.visualstudio.com/docs/copilot/chat/mcp-servers

(The above documentation is for VS Code 1.102, but Positron is currently using 1.100, so not all features are available. We're working on an update.)

[zachcp]

Beautiful. Will give it a proper go. Thank you.

[zachcp]

Can the Assistant serve HTML via the Viewer? Creating Plots is fantastic, generating HTML opens up the world.

[juliasilge]

Can you say a bit more about your use case? We are tracking a related feature request in #7324, although that may not be the kind of Viewer use you are meaning here.

[zachcp]

Hi @juliasilge , thank you for your great work.

The feature I am looking for is something more akin to Claude's Artifact - a single page HTML. I would like to explore using scientific javascript like 3dmol.js. I know RStudio used to have nice support for serving files single page files through methods like HTMLWidgets so i was hoping there might be a low-lift way to explore via chat and display in the HTML in the editor.

[juliasilge]

Ah OK, interesting 👀

Would you open an issue with your feature request, with some details of what you are looking to do? Thanks!

[kendavidn]

Great work deploying this, team 🙏. It's a huge unlock.

Agent mode was added recently: #8098. I tested it out and it worked great — the session context seems to give better results than the equivalent agent in Cursor.

One issue: I’d really like it to auto-run code without prompting me each time. Ideally, there’d be a setting to allow this. Cursor and Windsurf prompt for it after the first run request, with all the necessary warnings.

To prevent token overuse, maybe set a default limit — e.g. auto-stop after 10 tool calls. I believe Cursor does something like that.

[jmcphers]

You can enable code auto-run in this setting:

We'll consider the other features!

[CoryMcCartan]

Great to see this. I tried out the inline completions but had to turn them off because they were overriding the default autocomplete for file paths and function arguments. So, e.g. Copilot was suggesting a made-up file path and I was unable to tab/select through the actual files in my directory the way one normally can when one starts typing a file path like "~/dir/...

Is there any way to not hide the default autocomplete window that shows file paths and function arguments when using the inline completions? I think in RStudio the two can coexist

[juliasilge]

Thank you for this feedback! 🙌 I entirely agree, and I opened #8702 to track and prioritize improving this area.

A pretty bad workaround right now is to use Ctrl+Space to bring up regular completions, but you have to do that for every segment of the path to get the regular completions.

[matthewgson]

Can't wait to have a fully-featured Positron Assistant. I keep going back and forth from VSCode and Positron only due to the Agent mode. Are agent modes from both Antrophic and Copliot (and others) on the roadmap?

— thanks I will try!

[kendavidn]

Agent mode is currently live actually! #8098

You can enable code auto-run with the Chat > Tools > Auto Approve setting.

[juliasilge]

You can switch between our three modes (ask, edit, and agent) in this bit of UI:

If you do not have chat.tools.autoApprove enabled, you can still use agent mode but it will prompt you to approve each bit of code before it is run.

[matthewgson]

Oh, I guess it is available for Anthropic but not Github Copilot yet?

[juliasilge]

Correct; you can follow along for GitHub Copilot Chat support at #2010.

[ntluong95]

Can I ask how to change the completion provider for Copilot? Let's say gpt5 instead of Gemini 2.0 Flash

[jmcphers]

Do you have the Gemini Code Assist extension installed? It's possible you're seeing information from that extension; Github Copilot doesn't offer Flash as a model to the best of my knowledge. There is a setting github.copilot.selectedCompletionModel that provides what you're asking for in theory, but in practice Copilot only has one supported model. See here for more info: github/copilot-language-server-release#2

In any case I think you'd find GPT-5 too slow and expensive for use as a completion provider; we did some early experiments with using chat models (like GPT-5) to power the completion system and the results weren't very good. Github's Codex model (referenced in the above) is specifically tuned for low-latency completions code and probably gives you better results than GPT-5 would!

[jobreu]

Thanks a lot for this great feature! I activated the assistant and added GitHub Copilot via my account. However, when using the chat (inline or in the chat window), there is no response. It only shows the status "working...". I am using Windows 11. Is that related to the OS or am I maybe missing anything?

[jobreu]

@davidrsch Thanks a lot! I will definitely try this out!

[davidrsch]

Yeah, I was checking again and saw it actually released back in June, guess I’ve been extra distracted. 😅

[jmcphers]

This extension will not work as you expect in Positron because Positron Assistant is already using some of the internal chat APIs and some coordination is needed between the two systems. The next stable release of Positron (2025.10) will contain a bundled copy of the Copilot Chat extension that has been modified to work in Positron! (it's also in the dailies if you're feeling brave but be advised that it has not been thoroughly tested yet)

[davidrsch]

Thanks for the clarification! I was wondering, if I disable Positron Assistant, would that allow me test the Copilot Chat extension directly?

[jmcphers]

No, it won't. This may have changed, but last I checked, Microsoft's servers do not recognize Positron as a valid client of the Copilot API. So you may be able to sign in, but won't be able to use the service.

[davidrsch]

Reading about Databot got me thinking, it looks amazing, and it seems like it could be really powerful if integrated into Shiny apps alongside shinychat, maybe creating an experience similar to Looker Pro or Power BI, allowing users to dinamically create dashboards and report using natural language. I’m also curious if there are plans for Databot to support other models beyond Anthropic, I realize that might be better suited for a separate discussion, but it seemed worth mentioning here.

[jthomasmock]

Thanks!

Databot will have parity with Positron Assistant's access to model providers long-term, so as we add models to Assistant they will eventually also be available in Databot.

If you like the idea of Databot, but wanted to integrate into your own applications, check out querychat: https://github.com/posit-dev/querychat

Live Example: https://jcheng.shinyapps.io/sidebot/

It's NOT 1:1, but it does give some of the natural language, but still code-centric way of interacting with data on top of app or dashboarding tools like Shiny

[davidrsch]

Those are good news, and thanks for recomending querychat!!