Deephaven MCP

Most data tools force you to choose:fastorreal-time. With Deephaven's revolutionary live dataframes, you get both. Process streaming data at millisecond speeds while your AI assistant helps you build, query, and analyze — all through natural language.

- Live Data, Live Results: Query streaming Kafka, real-time feeds, and batch data as easily as static CSV files
- AI-Native Integration: Your AI assistant understands your data pipeline and can help optimize, debug, and extend it
- Enterprise Ready: Battle-tested on Wall Street for over a decade, now available for your team
- Zero Learning Curve: Write queries as if working with static tables — real-time updates happen automatically

It is built for data scientists, engineers, analysts, and business users alike — whatever your programming experience. Let AI generate the code while you focus on insights.

Deephaven MCP ships as a single package, and onedh-mcp-systems-serverreadsone configuration directory tree. That tree can hold acommunity/section, anenterprise/section, orboth at once— the single server hosts everything it finds, simultaneously. Each section is optional; you are never locked into one deployment type.

Install once (below), then configure whichever sections you need.

uv tool install --python-preference managed "deephaven-mcp"

About--python-preference managed: tellsuvto download and use its own managed Python (under~/.local/share/uv/python/) instead of any Python on your system. You do not need to install Python yourself.

uv tool install --python-preference managed mcp-proxy

One command writes a working configuration:

No prompts, nothing to edit by hand. You can now start a Deephaven worker whenever you want one — no Docker, nothing else to install:

Optional— to use a Deephaven server you already run, add it:

dhcli config session add local --host localhost --port 10000 \ --auth psk --token '${env:DH_LOCAL_PSK}'

The${env:...}form keeps your token out of the file, so set it in your shell:export DH_LOCAL_PSK='your-token'. Use--auth anonymousif your server needs no token.

To check your configuration at any time:

Exit code0means you are good to go.dhcli config filesshows where the files landed.

Confirm the configuration is good before involving an AI tool.dhclistarts its own background server, so there is nothing for you to launch:

dhcli config validate # is the configuration itself well-formed? dhcli session list # can the server load it and see your session?

config validatechecks the files alone.session listgoes further: it loads the tree and lists the sessions the server knows about. If your session appears, both the configuration and the server are working, and anything that goes wrong from here is in your AI-tool wiring. A bad configuration fails immediately with an error naming the file and field.

Your AI tool starts the systems server for you and shuts it down when it exits. There is no port to pick, no shared secret, and no background process to manage.

For Claude Desktop, openClaude Desktop→Settings→Developer→Edit Configand add:

The default config directory is~/.deephaven/ai/config/(created in Step 1), so noDH_AI_DATA_DIRis needed. To use a different location, add anenvblock settingDH_AI_DATA_DIRto a data root that contains aconfig/subdirectory.

Restart your AI tool (or IDE) so it picks up the new configuration.

"List my Deephaven sessions and show me the tables in the local session"

"What Python packages are installed in my Deephaven environment?"

"Execute this Python code in my Deephaven session:t = empty_table(100).update('x=i', 'y=i*2')"

One command declares your system. Each Enterprise system becomes one file underenterprise/systems/.

Password auth (with the secret read from an env var):

dhcli config system add prod \ --url https://dhe.example.com/iris/connection.json \ --auth password --username iris --password '${env:DH_PROD_PASSWORD}'

Private-key auth— PEM text is multi-line, so reference the file:

dhcli config system add prod \ --url https://dhe.example.com/iris/connection.json \ --auth private_key --key '${file:/etc/deephaven/prod-key.pem}'

Omit any flag to be prompted for it on a terminal. Then verify:

export DH_PROD_PASSWORD='your-password' # for the password-auth variant dhcli config validate

Export the secret before runningvalidate— it resolves${env:...}references in your current shell. The name you pass becomes the filename, soprodabove createsenterprise/systems/prod.json.communityis reserved and cannot be used as a system name.

Multiple systems: rundhcli config system add <name> ...once per system. List what you have declared withdhcli config system list, and remove one withdhcli config system remove <name>.

Community and Enterprise live side by side in the same tree — the one server hosts both:

~/.deephaven/ai/config/ ├── community/ │ └── sessions/ │ └── local.json # Community Core sessions └── enterprise/ └── systems/ └── prod.json # Enterprise systems — same tree, one server

Confirm the configuration is good before involving an AI tool.dhclistarts its own background server, so there is nothing for you to launch:

dhcli config validate # is the configuration itself well-formed? dhcli system list # which systems does the server serve?

system listreturns every configured system as{name, type}pairs — thecommunityumbrella alongside each Enterprise system. Yourprodentry appearing there confirms both the configuration and the server are working. A bad configuration fails immediately with an error naming the file and field.

To check that the system is actually reachable, not just configured:

dhcli system status --system prod --connect

Your AI tool starts the systems server for you and shuts it down when it exits. There is no port to pick, no shared secret, and no background process to manage.

For Claude Desktop, openClaude Desktop→Settings→Developer→Edit Configand add:

The default config directory is~/.deephaven/ai/config/(created in Step 1), so noDH_AI_DATA_DIRis needed. To use a different location, add anenvblock settingDH_AI_DATA_DIRto a data root that contains aconfig/subdirectory.

Restart your AI tool (or IDE) so it picks up the new configuration.

"What is the status of my Deephaven Enterprise system?"

"List all persistent queries on my enterprise system"

"Show me the tables available in my enterprise session"

Already havedeephaven-mcpinstalled?Here's how to upgrade:

uv pip install --upgrade "deephaven-mcp"

Using standardpip(venv-based install):

.venv/bin/pip install --upgrade "deephaven-mcp"

After upgrading, restart your AI tool for changes to take effect.

The Deephaven command-line tool, designed for humans and especially AI agents. It inspects and operates Deephaven systems from the shell with noun-verb commands, typed flags, and machine-first structured output.