SmitheryListed

Etincel

# Étincel Find the AI tells in your prose. Deterministically, locally, in your agent's own drafting loop. Étincel gives Claude, Cursor, or any MCP client two things most writing tools don't pair together: a trained voice to draft in, and a rule-based audit to check the result against. Neither step calls a model. The audit runs on pattern and rhythm analysis against your own past writing, not a wrapped API call to another LLM. ## Tools an agent gets - `audit_text`: scores a draft for AI-writing tells (uniform paragraph length, hedged authority, stock transitions), with locations and severities, not just a verdict - `train_style` / `create_style_from_dials`: builds a voice profile from your own writing samples or explicit parameters - `check_voice_match` / `check_self_repom a trained voice, and habits repeatedacross a writer's own past pieces - `get_style_guide` / `list_styles`: pur a saved voice before drafting starts - word-list and instruction tools to tune what counts as a tell per style ## Why this fits an agent's workflow, not just a human's A skill file that says "don't sound like AI" is a suggestion the model can silently ignore. `audit_text` returns a score an agent can branch on:audit, done. It's a gate, not a hint. ## Measured, not asserted Pooled AUC per register, checked againsixed bootstrap seed, in CI: | Register | AUC | |---|---| | memo | 0.909 | | essay | 0.900 | | blog | 0.820 | | docs | 0.735 | Email and general-register scores are tracked too but are deliberately uncalibrated; the four above are the ones to trust. ## Install \`\`\`json { "mcpServers": { "etincel-nonfiction": { "command": "npx", "args": ["etincel", "serve"] } } } \`\`\` No account, no network call in the audit path, MIT licensed.

First seen 2 Oct 2026. Evidence as of 5 Oct 2026.