KrillinAI CLI

Use this as the top-level routing skill for KrillinAI command-line work.

Build And Locate The Binary

Build both KrillinAI binaries through the repository script:

pnpm krillinai:build

For the current machine, locate the CLI without assuming a platform or architecture:

TARGET="$(node -p "process.platform + '-' + process.arch")"
SUFFIX="$(node -p "process.platform === 'win32' ? '.exe' : ''")"
KRILLINAI_CLI="$PWD/.runtime/build/krillinai/$TARGET/bin/krillinai-cli$SUFFIX"
test -f "$KRILLINAI_CLI"

The CLI loads config/config.toml relative to its process working directory. For real runs from a source checkout, use runtime/krillinai as the process working directory and create its ignored config/config.toml from config/config-example.toml without overwriting an existing file. A release archive should be run from its unpacked package root.

For build layout, configuration, JSON, manifest, output, and error details, read:

skills/krillinai-cli/references/cli-contract.md

Choose The Command

User intent	Command / skill
Generate source, target, bilingual, and short vertical subtitles	subtitle; use krillinai-subtitle
Generate target-language dubbing from subtitles	tts; use krillinai-tts
Create landscape videos	render-horizontal; use krillinai-render-horizontal
Create portrait/short-form videos	render-vertical; use krillinai-render-vertical
Generate a cover from a complete image prompt	cover; use krillinai-cover
Generate speech from plain text	speech
List provider voice codes for Aliyun, OpenAI, or MiniMax	voices
Validate a requested multi-stage output plan	pipeline --dry-run; use krillinai-pipeline

pipeline only validates plans in dry-run mode; non-dry-run execution returns unsupported_command. status is reserved and unsupported.

Operating Rules For Agents

• Use a dedicated --workdir; do not scatter outputs in the repo root.
• Use --dry-run for command-shape validation before external calls.
• Parse stdout as JSON lines. The terminal response is the object containing ok; OpenCreator mode may emit progress frames first.
• Treat krillinai_manifest.json and actual output files as the source of truth after a real stage.
• Reuse manifest outputs for later stages instead of guessing filenames.
• If a command fails, classify by error.kind: usage, retryable, dependency, or internal.
• Avoid rerunning expensive stages if the manifest already has valid upstream outputs.
• Do not claim that cover consumed a reference image: the current CLI accepts a complete text prompt and size only.
• Do not request an Edge TTS voice catalog: voices currently supports aliyun, openai, and minimax.

Minimal Workflow

REPO_ROOT="$PWD"
WORKDIR="$REPO_ROOT/tasks/demo"
mkdir -p "$WORKDIR"

(cd "$REPO_ROOT/runtime/krillinai" && "$KRILLINAI_CLI" subtitle \
  "https://www.youtube.com/watch?v=VIDEO_ID" \
  --origin-lang en \
  --target-lang zh_cn \
  --workdir "$WORKDIR" \
  --caption-source any \
  --prepare-video)

(cd "$REPO_ROOT/runtime/krillinai" && "$KRILLINAI_CLI" render-vertical \
  --workdir "$WORKDIR")