Frequently asked questions

This page collects the questions new users ask most often. If your question is about a specific client, the Coding Sessions pages and Getting Started go into more depth; for charges see Billing and Service Units; for failures see Troubleshooting.

Access and cost

Who runs this, and how do I get help?

The service is run by RCC staff. For access requests, problems the Troubleshooting page does not solve, or requests that need RCC staff (a new model, a larger fine-tuned adapter, a longer time limit), open a ticket through the standard RCC support channel and mention ai-session.

How do I get access?

You need an RCC account (no special group membership is required) and a Slurm account and GPU partition to run the GPU job under — the latter are unique to you and your PI, so you supply them the first time you start a session (--account and --partition, then remembered). Once you have those, you load the module and start a session as described in Getting Started. There is no separate signup step.

What will a session cost?

Sessions are billed in Service Units (SU), where 1 SU is one A100-GPU-hour. You are charged the greater of two things: the GPUs you hold (w_gpu × N × hours) or the tokens you process. For interactive work the hold cost almost always dominates. The default coding session (Qwen2.5-Coder-32B on two A100 cards) costs 2.0 SU per hour. The full formula and the measured per-model rates are in Billing and Service Units.

How do I keep the cost down?

Run ai-session stop the moment you stop working. An idle session still holds its GPUs and still bills. For interactive use the standard A100 configurations have the lowest hold cost; the faster H200 configuration only pays for itself under sustained, high-throughput generation.

Is my data private?

Your prompts and the model's completions are served entirely on the cluster and are not sent to any outside provider. Your browser-chat history is stored in a private directory in your home folder that other users cannot read. One caveat: coding tools can have their own telemetry that is separate from the model service. The service disables that telemetry where it controls the configuration, but you should confirm the settings in any client you install yourself. A second caveat, opt-in: if you start browser chat with AISESSION_TOOLS=1 (web search, URL fetch, reference lookup — see below), those specific tool requests send your query terms to services outside RCC. They are off unless you set that flag.

How do I connect from my laptop?

Open an SSH tunnel from your laptop to the login node your session is running on, then point your client at http://localhost:<port>. The exact tunnel command, with the right port and login node filled in, is printed when you run ai-session connect and is shown in Getting Started.

Yes. The person who starts the session receives an access key. Share that key with your labmates; each of them opens their own SSH tunnel to the same login node and uses the key as their API key. Anyone without the key is refused. All usage bills to the person who started the session, so coordinate within the group on who runs it.

Which model should I use?

Use Qwen2.5-Coder-32B for code, the Qwen2.5-72B general model for mixed prose-and-code work or when you specifically want the largest general model, and Qwen3-4B for quick or low-cost tasks. For math and multi-step planning, the Qwen3 thinking models reason before answering — Qwen3-32B (--model qwen3_32B) on two A100s, or the small Qwen3-4B. All served models are text-only; if you need a vision model (for images), ask RCC staff.

Whichever you end up on, start small and scale up: get your prompts or agent setup working against the small model first — it loads faster, spends less time waiting for free GPUs, and costs the least per hour — then switch to a larger model without changing any client configuration. If you need larger, ask; new models are staged on request. For a rough sense of how these open models compare to closed "frontier" models, see the capability frame of reference.

Can the browser chat search the web or find papers?

Yes, opt-in. Start browser chat with AISESSION_TOOLS=1 ai-session chat to add three tools you enable per conversation in Open WebUI: web search, URL fetch, and academic reference search (arXiv, bioRxiv, medRxiv, PubMed, Semantic Scholar). All three work with any served model, orchestrated by the UI. For the model to place the reference-tool calls itself, start a tool-calling model with the --agent flag — AISESSION_TOOLS=1 ai-session chat --model qwen3_32B --agent. These tools reach outside RCC (see Is my data private? and Getting Started); they are off by default.

Coding agents, MCP, and building agents

Which coding tool should I use?

aider is the dependable default for editing files and does not need tool-calling. opencode is supported for full tool-calling agents but requires a small workaround file (see opencode and Cline). Continue is the choice for in-editor use inside VS Code or JetBrains.

My agent said it made a change, but nothing happened. Why?

The Qwen2.5-Coder-32B model does not emit the tool-call markers that the model server's parser expects, so tool calls can fail silently: the agent reports success but no file is edited. Use the AGENTS.md workaround documented on the opencode page, or switch to the Qwen2.5-72B or Qwen3 model for agent work, which do not have this problem.

How do I add an MCP tool to my agent?

Add an mcp block to a project-local opencode.json in your working directory, not to your personal configuration; ai-session mcp config prints a ready-to-paste block for the two built-in tool servers (see MCP Servers). Before you enable a server, read the agent-responsibility guidance: an MCP server runs with your full cluster permissions and can reach any file you can reach, including a labmate's files through shared project directories.

Can I build my own agent on these models?

Yes. Point any agent framework that speaks the standard OpenAI API format — for example PydanticAI, LangGraph, smolagents, or the OpenAI Agents SDK — at the session URL, using your session key as the API key. For reliable tool-calling, use the Qwen2.5-72B or Qwen3 model rather than the Coder model.

Common problems

My session sits in the queue and never starts.

The GPUs you requested may be busy. A session waits for cards to free up, and the start command gives up if it waits too long. Try a smaller model (ai-session fast), or try again later.

My connection worked and then stopped.

If the login node was rebooted, or your SSH session dropped, the gateway — the connection point on the login node — can stop while the GPU session keeps running. Check with ai-session status; if the connection point is gone but a server is still listed, run ai-session stop to release the GPUs and start again.

My home directory filled up.

Browser-chat history is stored in your home directory, which has a smaller quota than project space. Clear old conversations in the chat interface, or remove old files under $HOME/.ai-session/.

I forgot to stop my session and was billed for idle time.

An unused session bills its GPUs until its time limit expires. Always run ai-session stop when you finish. If you routinely forget, ask RCC staff whether the idle-session reaper is enabled, which warns and then ends sessions that have gone quiet.