pangram

name	pangram
description	Use Pangram's MCP tools for bounded AI detection, plagiarism, combined text analysis, task and bulk inspection, and explicitly enabled local history or configuration work. Use when an agent needs Pangram analysis, task or bulk polling, canonical result interpretation, or a capability advertised during server discovery.

Pangram

Use the typed Pangram tools discovered from the server. Treat detection as probabilistic evidence, not a verdict about authorship.

Work safely

1. Discover the server before calling a tool. Use the advertised inventory; capability-gated tools can be absent.
2. Set a positive max_billable_units on every billable submission. Pangram 4 text costs one unit per started 100-word block, with a minimum of one. Bulk requests sum that rule across items and cannot exceed 1,000 units. Plagiarism uses 5 units. Combined analysis adds 5 to the text estimate.
3. Keep save and public_link false or omitted unless the user asked for retention or a public link and the server advertises the required gate. Sending content to Pangram is inherent to analysis; local retention and public links are separate choices and are disabled by default.
4. Read the canonical envelope from structuredContent. A domain failure has isError: true but still returns a complete canonical failure envelope. Use the text content only as a short summary.

Choose a tool

• Use detect_text for one text analysis. Supply text and the billing ceiling. Request input echo only when needed.
• Use check_plagiarism for one plagiarism-only text check. Its ceiling must allow 5 units.
• Use analyze_text for ordered AI detection followed by plagiarism. Its ceiling must allow the text estimate plus 5 units. A failed plagiarism check can return a partial analysis that preserves the successful AI result.
• Use get_task or wait_task with exactly one ID. An upstream_task_id works without local history; an analysis_id requires the history gate.
• Use submit_bulk with exactly one of inline items or jsonl_path, plus the billing ceiling. A JSONL path also requires an approved file root.
• Use get_bulk, wait_bulk, or get_bulk_results with exactly one local or upstream bulk ID. Local IDs require history. Results require explicit offset and limit; the limit is 1 through 1,000.
• Use check_update only when the user asks to check for an update. 0.x builds advertise the tool but return update_unavailable without network or state access.
• Use history and configuration tools only when server discovery advertises them. save: true requires both history and history-mutation capability.

Cancellation stops local observation only. It does not cancel a submitted Pangram task or bulk job. Do not report upstream cancellation.

Load exact references

• Read pangram://schema/output/v1 for the canonical result envelope.
• Read pangram://schema/errors/v1 for error and exit semantics.
• Read pangram://skills/pangram when the full embedded skill is needed.

Do not call or claim detect_files. Do not use the experimental MCP Tasks extension, invent prompts, or assume history resources exist.