MSK AI Portal and Databricks for opencode, Claude Code and omp

⚠️ All of this is API billing: you pay for every token. Be mindful of your token usage and costs.

Use opencode, Claude Code or omp with the models on:

• MSK's AI Portal (Open WebUI at https://chat.aicopilot.aws.mskcc.org), through a LiteLLM proxy that runs on your machine
• a Databricks workspace's Foundation Model APIs, directly, with the profile you already use for the databricks CLI (see Databricks)

aiportal-opencode                     # opencode with Claude Sonnet 5.5 on the AI Portal
aiportal-opencode -m gpt-5.6-terra    # or any other model on the portal
aiportal-claude                       # Claude Code on the AI Portal
aiportal-omp                          # omp on the AI Portal
databricks-opencode                   # opencode with Claude Sonnet 5.5 on Databricks
databricks-claude                     # Claude Code on Databricks
databricks-omp                        # omp on Databricks

Install

You need uv (curl -LsSf https://astral.sh/uv/install.sh | sh) and opencode, Claude Code or omp.

uv tool install git+https://github.com/msk-mind/msk-claude
aiportal setup

For the AI Portal, aiportal setup asks for your session token and checks the whole setup. To get the token: log in to https://chat.aicopilot.aws.mskcc.org, open Developer Tools (F12; Cmd+Option+I on a Mac), go to the site's cookies (Chrome/Edge: Application tab; Firefox/Safari: Storage tab) and copy the value of the token cookie.

Tokens last 14 days. When yours runs out, the launch commands stop and ask you to run aiportal token set with a fresh one; they also warn you 2 days ahead.

Update with uv tool upgrade aiportal, which also brings the latest model list. Uninstall with aiportal proxy stop && uv tool uninstall aiportal; your token and state stay in ~/.config/aiportal and ~/.local/state/aiportal until you delete them.

Coming from the Clone-Based Setup

If you used an earlier version from a clone of this repository, update the clone and run its installer once:

cd path/to/msk-claude && git pull && ./install.sh

It installs the package, replaces the old command links, stops the old proxy and moves your token and proxy key to their new place, so sessions you already have open keep working. It then lists the old files in the clone that you can delete. After that you don't need the clone.

Usage

opencode

aiportal-opencode                                  # default model, current directory
aiportal-opencode -m claude-opus-5.5               # a specific model
aiportal-opencode -m gpt-5.6-terra ~/project       # a model and a directory
aiportal-opencode --help                           # list models

Your global opencode config (~/.config/opencode/opencode.json) is never modified: the MSK provider is passed to opencode through OPENCODE_CONFIG, so your other providers and your default model stay as they are.

Claude Code

aiportal-claude
aiportal-claude -m claude-opus-5.5 ~/project
eval "$(aiportal env)" && claude      # or set up the environment and run claude yourself

omp

aiportal-omp
aiportal-omp -m gpt-5.6-terra ~/project

omp only reads custom providers from ~/.omp/agent/models.yml, so aiportal-omp adds an msk provider there (or updates it), and leaves the rest of the file as is. The proxy key isn't written to the file: omp reads it from the MSK_PROXY_KEY environment variable. The provider points at the proxy of the machine you last launched aiportal-omp on, so if machines share your home directory, start omp with aiportal-omp rather than plain omp.

Resuming Sessions

All three launchers can pick up an earlier session in the same directory:

	opencode	Claude Code	omp
Continue the most recent session	aiportal-opencode -c	aiportal-claude -c	aiportal-omp -c
Resume a specific session	aiportal-opencode -r <session-id>	aiportal-claude -r <session-id>	aiportal-omp -r <session-id>
Choose from a list	/sessions inside opencode	aiportal-claude -r	aiportal-omp -r
Continue in a copy, leaving the original as is	add --fork	add --fork	not available

With aiportal-claude, -m sets the model for the resumed conversation. Options and the directory can come in any order, e.g. aiportal-opencode ~/project -c.

Managing the Proxy and Models

You don't need to start the proxy yourself. aiportal-opencode, aiportal-claude and aiportal-omp check your token, reuse this machine's proxy if it's running, and otherwise start one in the background. A proxy keeps running after you exit, and is restarted automatically (on the same port, so open sessions keep working) when your token, its key or aiportal itself changes. Use aiportal for everything else:

Command	What it does
aiportal setup	First-time setup: token, moving over from the clone-based setup, checks
aiportal doctor	Check the whole setup and say what to fix
aiportal token	Show when your session token expires
aiportal token set	Save a new session token (checked with the portal first)
aiportal proxy status	Where the proxy is running and whether it answers
aiportal proxy logs	Follow this machine's proxy log
aiportal proxy stop	Stop this machine's background proxy
aiportal proxy start	Run the proxy in the foreground, to watch its output
aiportal models	List the configured models
aiportal models sync	Show models the portal has added or retired, compared with this version's list
aiportal models sync --test	Send a one-line prompt to every configured model
aiportal env	Print the environment for running claude yourself

Available Models

Claude Models:

• claude-sonnet-5.5 - Claude Sonnet 5.5 - 1,000,000 context / 128,000 max output
• claude-sonnet-5 - Claude Sonnet 5 - 1,000,000 context / 128,000 max output
• claude-sonnet-4.6 - Claude Sonnet 4.6 - 1,000,000 context / 64,000 max output
• claude-sonnet-4.5 - Claude Sonnet 4.5 - 200,000 context / 64,000 max output
• claude-opus-5.5 - Claude Opus 5.5 - 1,000,000 context / 128,000 max output
• claude-opus-5 - Claude Opus 5 - 1,000,000 context / 128,000 max output
• claude-opus-4.8 - Claude Opus 4.8 - 1,000,000 context / 128,000 max output
• claude-opus-4.6 - Claude Opus 4.6 - 1,000,000 context / 128,000 max output

OpenAI GPT Models (via Azure):

• gpt-5.6-terra - GPT-5.6 Terra - 922,000 context / 128,000 max output
• gpt-5.6-luna - GPT-5.6 Luna - 922,000 context / 128,000 max output
• gpt-5.5 - GPT-5.5 - 922,000 context / 128,000 max output
• gpt-5.4 - GPT-5.4 - 922,000 context / 128,000 max output
• gpt-5.2 - GPT-5.2 - 272,000 context / 128,000 max output
• gpt-5 - GPT-5 - 272,000 context / 128,000 max output
• o3 - o3 - 200,000 context / 100,000 max output
• gpt-4.1 - GPT-4.1 - 1,047,576 context / 32,768 max output
• o1 - o1 - 200,000 context / 100,000 max output

Other Models:

• amazon-nova-pro - Amazon Nova Pro - 300,000 context / 10,000 max output
• llama-3.3-70b - Llama 3.3 70B - 128,000 context / 8,192 max output
• glm-5 - GLM-5 - 200,000 context / 128,000 max output
• qwen3-vl-235b - Qwen3 VL 235B A22B - 262,144 context / 32,768 max output

The default is claude-sonnet-5.5. Haiku isn't available on the portal; Claude Code's Haiku requests go to Claude Sonnet 5.5.

Databricks

databricks-opencode, databricks-claude and databricks-omp use the Foundation Model API endpoints of a Databricks workspace. They need no proxy and no aiportal setup: the workspace and token come from a profile in ~/.databrickscfg, the one the databricks CLI uses (set one up with databricks auth login). Usage goes to the workspace's bill.

databricks-opencode                        # Claude Sonnet 5.5, profile DEFAULT
databricks-opencode -m gpt-6-sol ~/project # a model and a directory
databricks-claude -m claude-opus-5-5       # Claude Code (Claude models only)
databricks-omp -p dev                      # another profile, i.e. another workspace
databricks-opencode -c                     # continue the most recent session

The profile is DEFAULT unless you set DATABRICKS_CONFIG_PROFILE or pass -p. -c, -r and --fork work as for the AI Portal launchers (Resuming Sessions); omp has no --fork.

Model	Endpoint	Context / max output
claude-sonnet-5-5 (default)	databricks-claude-sonnet-5-5	1,000,000 / 128,000
claude-opus-5-5	databricks-claude-opus-5-5	1,000,000 / 128,000
claude-haiku-5-5	databricks-claude-haiku-5-5	1,000,000 / 128,000
gpt-6-sol	databricks-gpt-6-sol	922,000 / 128,000
gpt-5-6-terra	databricks-gpt-5-6-terra	922,000 / 128,000

The workspace offers more models (databricks serving-endpoints list), but these are the ones checked with tool calls in all three clients. Gemini and GLM endpoints don't work with opencode's tool calls.

How each client is set up:

• opencode: Claude goes through Databricks' Anthropic-format endpoint (/serving-endpoints/anthropic) and GPT through its OpenAI Responses endpoint (/serving-endpoints/responses). As with aiportal-opencode, the providers are passed in a separate config through OPENCODE_CONFIG, so your global opencode config isn't changed.
• Claude Code: ANTHROPIC_BASE_URL points at the Anthropic-format endpoint, and the opus, sonnet and haiku aliases map to the Databricks Claude models.
• omp: as with aiportal-omp, databricks-omp adds (or updates) two providers in ~/.omp/agent/models.yml, databricks-claude and databricks-openai, and leaves the rest of the file as is. They replace the Claude Code beta flags omp sends, which Databricks rejects, and set the Claude models to adaptive thinking. After the first launch you can also pick these models in plain omp, as long as DATABRICKS_FMAPI_TOKEN is set.

The token is passed in the DATABRICKS_FMAPI_TOKEN environment variable, so it isn't written to any config file. A profile with a personal access token works for sessions of any length. With an OAuth profile, the launchers get a token from databricks auth token, which lasts about an hour; after that, quit and continue the session with -c.

How It Works

opencode / Claude Code / omp ──> LiteLLM proxy (localhost) ──> MSK AI Portal (Open WebUI) ──> models

The proxy accepts OpenAI-format requests (opencode, omp) and Anthropic-format requests (Claude Code) and forwards them to the portal with your session token. Along the way it (aiportal/callbacks.py):

• removes parameters the portal doesn't support, and adds the chat_id field the portal requires
• puts a plain-language explanation in front of common errors, such as an expired token or a retired model

Where Things Live

~/.config/aiportal/
  auth.env              your session token
~/.local/state/aiportal/
  master_key            secret the launchers use to authenticate to the proxy
  <hostname>/           one directory per machine
    server_info         where the proxy listens, and what it was started with
    proxy.pid, proxy.log, proxy.lock
    opencode.json       opencode config generated at launch (databricks-opencode.json for Databricks)
~/.omp/agent/models.yml omp's own config, where the omp launchers add their providers

Both are readable only by you, and follow XDG_CONFIG_HOME / XDG_STATE_HOME (or AIPORTAL_CONFIG_DIR / AIPORTAL_STATE_DIR). Proxy state is kept per machine, so machines that share your home directory (on a cluster, for example) don't interfere with each other.

Sharing One Proxy Between Machines

By default the proxy only accepts connections from its own machine. To use one proxy from several machines that share your home directory, start it with MSK_PROXY_BIND=0.0.0.0 (for example MSK_PROXY_BIND=0.0.0.0 aiportal-opencode). Launches on other machines that have no proxy of their own then find it and use it, with the same key. Traffic between machines is plain HTTP.

Troubleshooting

Start with aiportal doctor. It checks the installation, your token, the model config, the proxy, the commands on your PATH and opencode / Claude Code, and prints the fix for anything wrong.

Errors starting with [MSK proxy] include the fix: run aiportal token set for a rejected token, or uv tool upgrade aiportal for a model the portal no longer offers.

"No connected db" (HTTP 400). The proxy rejected the session's key: LiteLLM reports a key it doesn't recognize this way. It happens when a session outlives the key it started with, for example after the key file was deleted and recreated. Quit and resume the session with aiportal-opencode -c or aiportal-claude -c. Ordinary proxy restarts don't cause this: the proxy comes back on the same port with the same key.

The proxy won't start. The launch command prints the end of the log. Run aiportal proxy start to see the full output in the foreground.

Claude Code hangs on the "trust this folder" prompt the first time you use it in a directory. Run plain claude there once to answer the prompt, then use aiportal-claude.

Security

• Your token and the proxy key are readable only by you; the launchers tighten their permissions if they loosen.
• The proxy listens on localhost and requires the key, so other users can't spend your token through it.
• SSL verification is disabled only for the connection to the portal, which uses an internal certificate.

Development

git clone https://github.com/msk-mind/msk-claude && cd msk-claude
uv run aiportal-opencode        # run from the source tree
uv run aiportal doctor

The code is in src/aiportal/: launch.py (the three AI Portal launchers), omp.py (omp's models.yml), databricks.py (the three Databricks launchers), proxy.py (starting and finding the proxy), server.py (runs LiteLLM), callbacks.py (request fixes and error hints), auth.py, models.py, doctor.py, migrate.py (aiportal setup) and cli.py. Dependencies are pinned to tested versions in pyproject.toml, because the proxy relies on LiteLLM internals; upgrade them deliberately and re-test with uv run aiportal models sync --test.

Updating the Model List

src/aiportal/litellm_config.yaml is the single source of the model list; the launchers, the opencode config and the list above are generated from it. When the portal adds or retires models:

uv run aiportal models sync --apply   # update the config and this README
uv run aiportal models sync --test    # check every model answers

Then bump version in pyproject.toml, commit and open a PR. Once it's merged, everyone gets the new list with uv tool upgrade aiportal.