eddy

name	eddy
description	Use when the user wants to turn raw video footage into a finished YouTube launch kit (a long-form edit plus vertical shorts, titles, description, chapters, and thumbnails) with Eddy, the local-first agentic video editor. Covers starting and monitoring edits, mining shorts, inspecting runs, and reading results through the eddy-mcp tools.

Driving Eddy

Eddy turns raw footage into a complete launch kit, entirely on the user's machine. You drive it through the eddy-mcp MCP tools. Eddy never publishes or uploads — it only writes local files under the run directory.

Long operations are jobs (start → poll → read)

A full edit takes 5–15 minutes, so the long tools are asynchronous. Never block on them:

1. Start — use eddy_edit_start(source, focus?, template?, format?, language?, repair?, dry_run?) for the default promise-level flow. It returns { job_id, run_dir } immediately. The job_id is the run's slug. Use eddy_run_start only when the user asks for lower-level run controls.
2. Poll — eddy_job_status(job_id) returns running | completed | failed | interrupted plus the current phase (transcribe, iteration_N, final_render, shorts, package, done). Poll every ~30s and keep the user informed; on failed the payload includes a log_tail.
3. Read — eddy_artifacts(run) returns the candidate titles, description, chapters, shorts ledger, and the final video path. eddy_run_inspect(run) returns raw run state + a final/ inventory.

The same start→poll→read pattern applies to eddy_run_start (lower-level full run), eddy_shorts_start (vertical clips only, no long edit), eddy_transcribe_start, eddy_render_start, and eddy_batch_start (a directory of sources).

Instant reads (no job)

• eddy_runs — every run with phase + best iteration.
• eddy_run_inspect(run) — one run's state and artifact inventory.
• eddy_doctor — hardware + provider detection and an environment preflight.
• eddy_profiles — configured per-channel run profiles.
• eddy_qa(run, iteration?) — deterministic QA (+ judge if a proxy exists).
• eddy_pick(run) — deterministic A/B title + thumbnail pick.

Rules

• Never publish or upload. Eddy produces local files only; uploading is out of scope.
• Destructive tools need confirmation. eddy_clean (reclaims scratch) and eddy_purge (deletes PII; full=true erases the whole run) refuse unless confirm=true. Always confirm with the user first; eddy_purge is irreversible.
• On-device runs — pass local_only=true to keep everything on the machine (no cloud brain, no model downloads, no cloud thumbnail APIs).
• Identify a run by its slug (the run directory name); use it as both job_id and run.
• Tutorials/lessons — pass format="tutorial" (or lesson/longform/podcast) to lift the length ceiling.