Invoke one named tool from the platform's agent tool registry directly, with no model in the loop — you choose the arguments. This is how an external agent uses Talonic's retrieval and provenance while driving control flow itself: query_data for a read-only SQL SELECT over the extracted data, describe_data for the queryable field list, get_document_markdown to read a document's text. Denied capabilities come back as an error naming the capability required; the platform re-checks every call. API-key credentials are restricted by the platform to the registry's read-only tools (capability data.read); write-capable registry tools are never invocable through this credential, so through an API-key credential this tool reads and never mutates workspace data.
Not listed on ChatGPT connections: OpenAI's plugin review requires every model-callable operation to be its own tool, so the hosted server omits talonic_list_agent_tools and talonic_invoke_agent_tool for ChatGPT. Every other client lists both.
When to use
talonic_list_agent_tools showed a tool with can_invoke: true that does what you need.
When not to use
Anything a dedicated talonic_* tool already does — prefer those.
Parameters
Parameter
Type
Description
name
string
Tool name, e.g. `query_data`.
args
object
Arguments matching the tool's `input_schema`.
document_ids
string[]
Hard scope for scope-aware tools.
Response shape
Fields
Parameter
Type
Description
tool
string
The tool that ran.
result
any
The tool's parsed output.
citations / artifacts / cards
array
Present when the tool attached them.
Tool input
{ "name": "query_data", "args": { "sql": "SELECT field_key, count(*) FROM cells GROUP BY 1 ORDER BY 2 DESC LIMIT 5" } }