Using Cursor

Pipemason is a dual-provider pipeline. By default it runs your agents on the Claude Agent SDK, but it can run the exact same pipeline on Cursor's cursor-agent CLI instead. The orchestration, phases, gates, and contracts are identical — only the agent backend changes.

Choosing a provider

Select the backend with the --provider flag (default claude). It is accepted by pipemason start, pipemason program start, and pipemason init:

# one-off: run a single story on Cursor
pipemason start "add a /version endpoint" --provider cursor

To make Cursor the default for a repo, set it once in .pipeline/config.yml (or scaffold it with init, below):

# .pipeline/config.yml
provider: cursor

A --provider flag on an individual run overrides the config value; with neither set, pipemason uses claude.

Prerequisites

The Cursor path needs two things on the machine running the pipeline:

• The Cursor CLI — the cursor-agent binary must be on your PATH.
• A Cursor API key — exported as CURSOR_API_KEY. Unlike the Claude path (which uses your Anthropic key), all inference and billing on the Cursor path run through your Cursor account.

export CURSOR_API_KEY=...   # from your Cursor account
pipemason doctor            # verifies cursor-agent is on PATH and the key is set

Scaffold with pipemason init

pipemason init --provider cursor writes provider: cursor into .pipeline/config.yml and scaffolds Cursor-native config so the agents run with the right roster and rules:

• .cursor/agents/*.md — the pipemason subagent roster
• .cursor/rules/*.mdc — pipemason's working rules
• .cursor/mcp.json — MCP server configuration

pipemason init --provider cursor

Models and cost

Pipemason maps its model tiers to a Cursor model id automatically; you do not pick raw model names. Inference and billing happen in your Cursor account, so there is no separate Anthropic bill on this path.

Because the cursor-agent CLI does not report per-spawn cost, spend on the Cursor path is estimated: pipemason counts tokens from the stream and prices them with its own table. The cost ledger and ceilings still work, but every figure is labelled "estimated" rather than billed-exact. See Cost control for how ceilings behave.

Limitations on the Cursor path

• Cost is estimated, not exact — use the Claude path if you need billed-accurate spend reconciliation.
• Native budget caps are not enforced — maxTurns, maxBudgetUsd, and maxThinkingTokens have no Cursor equivalent. Pipemason's own ceilings still apply (against estimated spend).
• Tool-call output is parsed defensively — the Cursor CLI's JSON schema is less stable than the Claude SDK's.