version: "1.0.0" name: typeagent-macros description: "Discover, inspect, validate, and run TypeAgent tool-composed macros. Use when a user asks to repeat a recorded workflow or run, adapt, repair, or manage a TypeAgent macro."

TypeAgent Macros

Use the typeagent-macros MCP server for macro catalog and lifecycle work.

Run A Macro

1. Use search_macros or list_macros to find the macro.
2. Use inspect_macro and get_macro_requirements before execution. Collect

all required inputs without exposing secret values in chat. Inputs may fill nested argument-template fields, and later steps may bind values from earlier results.

1. Call run_macro with preference: "auto".
2. For completed, report the sanitized result. For failed or cancelled,

report the structured failure without inventing a repair. A postconditionFailed result means the live tool output no longer has the shape captured by the approved procedure.

1. For agentRequired, invoke the TypeAgent Macro Runner agent with the

complete returned launch object. Do not manually paraphrase or reconstruct the launch payload.

Deterministic replay and agent-guided execution are whole-macro choices. Never replay a prefix before handing the remaining steps to the runner.

Adaptation

The runner may submit a changed successful procedure through submit_macro_candidate. Candidate submission creates a separately reviewable draft. It never changes or approves the source version. Permission denial, cancellation, and timeout are terminal outcomes, not adaptation signals.

Lifecycle

• Create drafts only from explicitly captured traces.
• Validate drafts before asking the user to approve them.
• Selected completed recordings automatically enter recipe extraction, generalized

building, grammar validation, and registration. Do not rerun the task to teach it.

• The workspace preference is Off, Prepare, Read-only, or All. Prepare requires

review; Read-only auto-approves read-only procedures; All auto-approves valid procedures including writes. Never change that preference for the user.

• Newly learned macros execute through the live macro runner. Macro approval does

not grant tool permission; denial and cancellation remain terminal.

• Use get_macro_run only for sanitized persisted deterministic run evidence.