Short answer
- Choose hosted MCP for the quickest Claude connection, minimal local setup, and managed remote tools.
- Choose a local skill pack for version-controlled workflows, local outputs, editable scripts, and richer pipelines.
- Use both only with separate responsibilities. A retry in one path must not recreate a paid job started by the other.
What is the official Arcads MCP route?
Arcads documents a hosted MCP server at https://mcp.arcads.ai. In Claude Desktop, the operator adds that URL under Settings → Connectors and completes the Arcads account connection. The connector is the tool transport and execution boundary; an active plan and remaining credits are still required for generation.
- Best fit: a fast managed connection in a supported MCP client.
- Advantage: fewer local runtimes, scripts, and dependency checks.
- Tradeoff: less direct control over local orchestration and custom script behavior.
Start with the official Arcads MCP documentation, then list available tools without generating anything. Approve one small request, record its job reference, and verify that a replay does not create another paid job.
What is a Claude Code or Cursor skill pack?
A repository-first skill pack supplies local instructions, scripts, prompt libraries, and workspace conventions that call the Arcads API or MCP tools. The open-source Arcads AI Video agent skill pack maintained by Caleb Kruse is a concrete example: its documentation covers a local API key, Python workflows, prompt libraries, output folders, per-call logs, and optional media-processing dependencies.
- Best fit: teams that want editable workflows under version control.
- Advantage: local artifacts, scripts, prompts, audit logs, and elaborate creative pipelines.
- Tradeoff: the operator owns dependencies, secret handling, updates, and script-level controls.
Where do official Arcads skills fit?
Arcads also publishes an official skills repository for guided commands and workflows. A skill is an instruction layer. It does not replace the hosted MCP URL, the account connection, an active subscription, or a credit limit. Write down which component may create jobs and which components may only prepare, validate, or poll them.
Decision table
| Requirement | Hosted MCP | Local skill pack |
|---|---|---|
| Fast Claude setup | Strong fit | More setup |
| Minimal local dependencies | Strong fit | Operator-owned |
| Cursor support | Client-dependent | Strong fit |
| Editable scripts and prompts | Limited | Strong fit |
| Local output organization | Client-dependent | Strong fit |
| Script-level audit logs | Client-dependent | Strong fit |
Five controls both routes need
- Approval payload: product, assets, script, model, quantity, and maximum credit consumption.
- One request ID: reuse it through creation, polling, and result delivery.
- Batch ceiling: reject requests above the approved quantity before a write call.
- Bounded polling: status checks may retry; job creation must not retry blindly.
- Fail closed: stop when approval, credentials, plan status, credit reading, or an asset is missing.
Before the first paid generation
Calculate the boundary with live plan values.
Check the approved batch, replacement allowance, reserve, and remaining headroom. Values stay in your browser.
FlowPatch Patch Pilot
Turn the selected route into one working workflow.
One fixed-scope Arcads setup with an explicit approval and credit boundary. The first $149.50 is due only after written scope acceptance.
FlowPatch is independent and is not affiliated with or endorsed by Arcads, Anthropic, Claude, Cursor, Caleb Kruse, or Kruse Media. Verify current setup, model, and billing details in the linked primary sources.