Skip to content

Reference

Troubleshooting and FAQ

Fixes for the failures people actually hit, plus straight answers about openness, offline use, data storage, and the remote endpoint.

2 min read · Updated

Troubleshooting

ProblemFix
Server does not startnode --version must be 18+. The server prints a clear stderr message when fetch is missing from the runtime.
Tools do not appear in the clientRestart the client fully after config changes; most clients only read MCP config at startup. Then check the config file path against Quick start.
scope "mine" requires an API keySet PASTEPILE_API_KEY in the server's env block and restart. Verify the key was not revoked at /keys.
Saves expire after a monthFree-key policy for defaulted expiry (expiry_note says so). Pass an explicit shorter expiry, or use a Pro key for never.
rate_limited responsesBudget exhausted; wait for Retry-After. Separate keys per assistant keep budgets independent. Numbers in Limits.
Cannot read a specific pasteIf the error names encrypted, password_protected, burn_after_read, or sealed, that is deliberate protection, not a bug. See Errors.

FAQ

Is the server open source and auditable?

Yes. The npm package ships readable, unbuilt source: one transport file and one pure dispatcher, zero dependencies.

Does it work offline?

The protocol layer does, but every tool call is an HTTPS request to Pastepile, so saving and retrieving need a connection.

What data does the server itself store?

Nothing. It is stateless and writes nothing to disk; all persistence is your pastes.

Can I point it at a self-hosted Pastepile?

Yes: set PASTEPILE_API to your deployment's base URL. See Configuration.

Is there a remote (HTTP or SSE) endpoint?

Yes. The hosted remote endpoint speaks MCP's Streamable HTTP transport and runs the same four tools as the stdio server. Tool calls on it require a free API key; the stdio server remains the anonymous option.

How is this different from the REST API?

Same capabilities, different consumer. The REST API serves your scripts; the MCP server packages it for assistants, with model-readable descriptions, memory-tuned defaults, and model-visible errors. The comparison table is in How it works.

Related documentation