AI SDK 3: Search and Execute Across Accounts

In toolMode: 'search_execute', @stackone/ai 3.1 gives the model one stackone_search_actions tool and one stackone_execute_action tool, plus stackone_list_accounts, across every account in scope. Before this release, each account added its own pair. stackone-ai 3.0 for Python runs every tool over MCP and is a breaking release, with a migration guide in the repository.

What’s new

• One pair for every account. The model searches and executes across all accounts in scope with the same two tools, so the tool list keeps its size as accounts are added.
• Accounts on search hits. Each search result carries available_accounts, so the model knows which accounts can run an action before it calls one.
• Ambiguous calls refused. When an execute call is ambiguous about which account should run it, the server refuses it and the SDK throws StackOneAPIError instead of picking one.
• Same names with one account. With a single account in scope, its pair keeps the same tool names, so prompts work the same either way.
• A 50-account ceiling. More than 50 accounts in scope raises a configuration error.
• Python on MCP. stackone-ai 3.0 executes every tool through MCP tools/call, adds tool_mode="search_execute" and StackOneToolSet.execute(action_id, arguments=None, *, account_ids=None, session_id=None), and retries HTTP 429 responses after the Retry-After delay.
• Python breaking changes. The package needs Python 3.11 or later, ships mcp as a core dependency in place of the [mcp] extra, moves LangChain support to the stackone-ai[langchain] extra, drops client-side search and returns results under result["result"]["data"].

Install

npm install @stackone/ai@^3.1.0
pip install "stackone-ai>=3.0.0"

Search and Execute introduced the two meta-tools these releases extend. The source lives in the Node and Python repositories, and stackone-ai 3.0.0 on PyPI lists the release.