toolstream

[](https://github.com/tylerwilliamwick/toolstream/actions/workflows/ci.yml) [](LICENSE) [](https://nodejs.org)

Every time Claude Code starts a conversation, it loads the full list of tools from every connected service. If you have GitHub, Jira, Confluence, and a few other services connected, that can mean 100+ tool definitions sent to the model on every single turn, costing tens of thousands of tokens before you've typed a word.

ToolStream fixes this. It sits between Claude Code and your services, and instead of loading everything upfront, it figures out which tools are relevant based on what you're talking about. If you're discussing a Jira ticket, the Jira tools appear. If you're working with files, the file tools appear. Everything else stays out of the way.

Result: 90%+ fewer tokens spent on tool definitions, with no loss of capability.

1. Claude Code connects to ToolStream instead of connecting to each service separately
2. ToolStream starts with just 4 small tools: discover_servers, discover_tools, execute_tool, and reconnect_server
3. As the conversation develops, ToolStream automatically brings in the tools that match what you're doing
4. If a tool doesn't appear on its own, Claude can search for it directly

flowchart TD
    Client["Claude Code"]
    TS["ToolStream\n─────────────────\nRoutes the right tools\nto the right conversation"]
    GH["GitHub\n(26 tools)"]
    AT["Atlassian\n(Jira + Confluence, 72 tools)"]
    OB["Obsidian\n(14 tools)"]
    OT["Other services..."]
    NT["Notifications\n(optional)"]

    Client -- "4 tools visible to Claude" --> TS
    TS -- "connects to" --> GH
    TS -- "connects to (with credentials)" --> AT
    TS -- "connects to" --> OB
    TS -- "co