export CLIPWRIGHT_API_KEY=cw_...

npx -y -p @clipwright/cli clipwright quote --script "Hello from Clipwright."
npx -y -p @clipwright/cli clipwright make  --script "Hello from Clipwright."

The package installs the same program under two names, clipwright and clipwright-cli. An unrelated npm package also called clipwright ships a binary under the short name, so installing both in one project leaves only one of them in node_modules/.bin — silently. Use clipwright-cli where that matters. The commands above name the package with -p, so they resolve here either way.

Recipes, each a prompt an agent turned into a video with the call it sent and the price, are in the @clipwright/mcp-server README.

Command	What it does
`quote --script <text>`	Estimates the credit cost. Spends nothing.
`make --script <text>`	Renders a clip and waits for it. Costs credits. Prints the video URL to stdout, progress and warnings to stderr.
`upload <path>`	Uploads a local PNG/JPEG and prints the https url to pass as `--image`. Free.
`runs <id>`	Prints the state of one run.
`voices`	Lists voices for `--voice` on `make`: the presets, the voices of `gemini-3.8-flash-tts`, then the catalog. `--skill make_faceless` lists the narration voices for `make-faceless` instead. Narrow it with `--language`, `--gender`, `--age`, `--use-case` and `--model`. Free.
`actors`	Lists the actors this account can use: the Clipwright catalog and your own. Free.
`quote-actor`	Estimates the cost of a personal actor. Spends nothing.
`create-actor`	Creates a personal actor from a description. Costs credits — one charge per requested format. Prints the run; follow it with `runs <id>`.
`quote-faceless`	Estimates the cost of a faceless video. Spends nothing.
`make-faceless`	Starts a faceless video: narration over an opening animated clip and image scenes. Costs credits. Prints the run; follow it with `runs <id>`.
`delete-actor <id>`	Deletes a personal actor of this account. Videos already made with it stay as they are. Free.
`shots <run>`	Lists the shots of a finished video: span, kind, start of the description and preview link of each. Free.
`edit-show <run>`	Prints the draft of an edit of a run as JSON. Version 0 means nothing is stored. Free.
`edit-save <run> --file <path>`	Saves the draft from a file `edit-show` printed. Charges nothing and renders nothing.
`revise-quote <run>`	Prices a revision of a faceless video and prints the quote as JSON. Spends nothing.
`revise <run>`	Starts a revision: a new run that renders the video again with the shots you change. Costs credits. Prints the run; follow it with `runs <id>`.
`candidate-quote <run>`	Prices a shot candidate, one picture or clip for one shot, and prints the quote as JSON. Spends nothing.
`candidate <run>`	Prints the quote of a shot candidate; with `--yes` makes it. Costs credits. Prints the run; follow it with `runs <id>`.
`candidates <run>`	Lists the shot candidates of a video's family as JSON, with state, charge and preview. Free.
`uploads`	Lists your uploads that a revision still takes as JSON, newest first, each with a fresh url and `usable_until`. Free.
`workflows`	Public workflow methods and availability as JSON. Free.
`workflow-show <method>`	One method's rules, requirements and limitations as JSON. Free.
`plan-create --file <path>`	Store `{ document }` without starting a run. Free.
`plan-show <plan>`	Print the current owned plan and revision as JSON. Free.
`plan-save <plan> --file <path>`	Save `{ base_version, document }` or the complete plan-show output for this plan. Free.
`plan-approve <plan> --file <path>`	Explicitly approve `{ revision, content_digest }`. Free; starts no run.
`plan-quote <plan>`	Print the approved plan's quote as JSON. Free.
`plan-run <plan> --file <path>`	Start one explicitly keyed paid workflow and print its receipt and native run immediately.

Workflow files and explicit purchases

The four native methods are ugc-product-intro, ugc-faq-answer, faceless-explainer and product-image-story, subject to their executor's configuration, qualification and activation. website-proof-demo and kinetic-launch remain execution_blocked with reasons.

Workflow commands print JSON to stdout; API refusal messages, including all issues and warnings, go to stderr. Files must be valid UTF-8 JSON and at most 1 MiB before parsing. Documents have a separate 256000 JSON-byte limit. Unknown write fields are refused; accepted raw narration is preserved.

Start with a method's complete example, then edit its brief and evidence to match your intent. Preserve the method/capability pins and accepted raw script:

clipwright workflow-show faceless-explainer > method.json
jq '{document: {v: 1, method: {method_id: .descriptor.id, version: .descriptor.version, content_digest: .descriptor.content_digest, capability_digest: .capabilities.content_digest}, brief: .descriptor.example.brief, beats: .descriptor.example.beats, claims: [], assets: []}}' method.json > create.json

Inspect the plan's readiness issues and warnings before choosing its exact revision for approval:

clipwright plan-create --file create.json > plan.json
jq '{revision, content_digest}' plan.json > approval.json
clipwright plan-approve wfp_example --file approval.json
clipwright plan-quote wfp_example > quote.json

Replace wfp_example with the returned plan id. After accepting the quoted price, choose one stable key for this purchase and extract exactly the launch fields. Do not send the whole quote as the launch file:

jq '{revision, content_digest, semantic_input_digest, quote_id, idempotency_key: "workflow-example:1"}' quote.json > launch.json
clipwright plan-run wfp_example --file launch.json
clipwright runs run_example

Use the returned run id in runs. The launch does not wait for the video. If its answer is uncertain, keep and explicitly resend the original file and key to recover that purchase; a new key can pay twice. No workflow command automatically approves, quotes, launches or selects a new attempt.

plan-show wfp_example > edited.json produces a file you can edit and pass to plan-save wfp_example --file edited.json. The CLI requires its original plan id, takes its revision as the base version, and validates the original document strictly. Future document fields survive reading and cause an unsupported save to fail before fetch. A repeated explicit save after a lost answer returns the already-saved current revision; a different current document conflicts.

runs appends optional workflow JSON with the purchased plan and available artifact/placement evidence. It keeps native state and warnings. Missing proof is explicit null with a validation issue; [] means verified no placements. Historical evidence does not renew a download URL, and creative_acceptance: "not_reviewed" is separate from technical delivery. Original product images use owned upload ids, digests and nonoverlapping word/quote anchors in supported IMAGE shots; the generated opener is illustrative. Required unsupported layouts or unproven protected placement are refused. See the workflow API contract.

Native video commands

quote and make take the actor's photo as --image <url> or --image-file <path> (uploaded first, then passed on), plus --aspect-ratio 9:16|1:1|16:9 and --resolution 720p|1080p|4k. Pass --aspect-ratio explicitly with a picture: without it a landscape photo renders vertical, with a warning.

Pick a voice with --voice <name> from that catalog, or --voice-id <id> for a vendor voice outside it (a cloned one, for instance). Set neither and the voice follows the actor's gender: sarah for a woman, george for a man. The gender comes from the --actor-id catalog entry, or from --actor-gender female|male next to --image. The default actor speaks george. --image without --actor-gender also gets george, and the response warns about it. --actor-gender with --actor-id is refused. voices prints a sample: link under each voice: play it before you pay for a render, and reuse the same name next time. --language filters by a voice's native language; any voice speaks any supported language.

The voice decides the speech model: every preset, catalog voice and --voice-id voice speaks eleven_v4. --tts-model overrides the choice: eleven_flash_v2_5 and eleven_turbo_v2_5 cost less but misread Russian stress marks. --tts-model eleven_v3 is still accepted, but the run speaks eleven_v4 and warns about it. In a Russian script, mark stress with a capital vowel inside the word: потОм. eleven_v4 and gemini-3.8-flash-tts read it as stress. eleven_v4 and gemini-3.8-flash-tts give no pause for a <break time="…" /> tag: the tag is cut from the text, and the response warns about it.

gemini-3.8-flash-tts speaks only its own thirty voices: voices --model gemini-3.8-flash-tts lists them. Naming one (--voice kore) selects that model without --tts-model; --tts-model gemini-3.8-flash-tts without --voice speaks kore for a woman and charon otherwise. A preset, a catalog voice, --voice-id and --captions are refused with it before any charge. Its voices have a sample in English, and voices --language ru links the Russian one.

quote-faceless and make-faceless take the narration as --script <text>, --script-file <path> or --brief <text> (a short description the narration is written from), plus a required --duration <seconds> from 30 to 90. The video ends with the narration, runs at least 25 seconds and may exceed the selection by up to a quarter, never beyond 90 seconds; a video more than 5 seconds longer says so in warnings[]. A script is accepted when its estimated length is 21 seconds to the selection plus 5, at most 90. The charge follows the delivered duration and never exceeds the quote; the rates and the minimum are on clipwright.io/pricing. Captions are on; --no-captions turns them off. --style-reference <url> and --character-reference <url> take public https images. --voice <name> picks the narration voice: take a name from voices --skill make_faceless, where each line shows the voice's language and gender. Without it the narration is read by narrator_en_wise_lady. A make voice such as george and an unknown name are refused before any charge. Quote first: both commands take the same flags, so the quote prices exactly the request make-faceless would send.

shots <run> prints text: the line of the run, whether it can be edited and until when its media are stored, then one line per shot with a preview link under it. The link points at a large source file; the command prints it and downloads nothing. A description or a link the API did not give is printed as withheld or none. A run with no shots to show prints the reason and exits with code 0.

edit-show <run> writes only the draft's JSON to stdout, so you can keep it in a file, change the document and pass the file to edit-save; the version and the issues of the draft go to stderr. edit-save <run> --file <path> reads version and document from that file, ignores its other keys and sends them as the base version and the document. A file of another shape is refused before any request. When the save is refused because the draft has another version, the command reads the draft once. If the draft already holds your document one version later, an earlier send went through and the save is reported as done. Otherwise it reports the conflict with the stored version: run edit-show again and apply your change to that draft. Write the new draft to another file than the one you pass with --file.

revise-quote and revise take one of two inputs: --shot <id> with one of --retake, --new-picture <description> or --animate [motion], or --document <path>, a file edit-show printed whose edited document is the whole revision. A revision that delivers is charged its quoted total; one that fails is charged nothing, even when its paid work was already done. The same request sent again returns the run it started; --retry 2 starts a second revision. While paid edits are closed on the server, both commands answer editing_disabled.

candidate-quote and candidate take --shot <id> with one of --retake, --new-picture <description> or --animate [motion], and --take <n>, 1 by default. candidate prints the price first and starts nothing without --yes. A delivered candidate is charged its one unit whether it is kept or discarded; one that fails is charged nothing, even when its paid work was already done. The same request and take sent again return the run they started; --take 2 makes another candidate, charged again. While shot candidates are closed on the server, the three commands answer editing_disabled.

make retries are explicit: --retry 2 deliberately starts a new run for the same script. Without it a repeated call returns the run you already paid for, because every request carries an idempotency key derived from the input and from this installation's id. A non-numeric --retry is rejected rather than coerced: the value ends up in that key, so a silent NaN would buy you a second render.

Set CLIPWRIGHT_CLIENT_ID to a stable per-install value, or leave ~/.clipwright/ writable so the client can create one on first use. It refuses to invent a throwaway id, because a per-process id would disable duplicate protection silently — and the duplicate it fails to catch is a second paid render.

CLIPWRIGHT_API_URL overrides the API base URL; the default points at production.