vivideo-toolchain

Vivideo Toolchain

The developer + agent integration layer for the Vivideo API — a shared client, an MCP server, a CLI, and Skills. Agents are a primary user, not an afterthought.

Everything sits on top of the public API (the source of truth). No backend business logic is duplicated here — the toolchain only adds client-side ergonomics and guardrails.

vivideo-toolchain/            npm workspaces monorepo (TypeScript)
├── packages/core   @vivideo/core   one typed API client + guardrails (shared)
├── packages/mcp    @vivideo/mcp    MCP server (stdio) — 13 agent tools
├── packages/cli    @vivideo/cli    `vivideo` CLI — humans, scripts, CI, agents
├── skills/                          reusable agent Skills (SKILL.md)
└── docs/                            installation, configuration, workflows

Architecture

One client, one set of types, derived from packages/core/openapi.yaml. The MCP server and CLI both call @vivideo/core — there is exactly one representation of each request/response, so the tools, CLI, docs and OpenAPI stay consistent.

@vivideo/core adds only client-side concerns, never business logic:

Guardrail	What it does
Rate limiting	Token-bucket cap on request rate (default 8/s) — the client can't become a request flood.
Concurrency cap	Semaphore limits in-flight requests (default 4).
Bounded retries	Retries only retryable failures, with full-jitter backoff, capped attempts.
Honors Retry-After	On 429 it waits the API-specified delay before retrying.
Idempotency	Every video-create sends an Idempotency-Key; retries reuse it → no duplicate charges.
Timeouts	Per-request AbortController timeout (default 30s).
Safe waiting	waitForVideo polls at the API's suggested interval, with a hard timeout — never a while(true).
Secret redaction	API keys / signing secrets are stripped from every log, error, and output.

These complement — and never bypass — the API's own auth, rate limits, idempotency, credit checks, premium gates, ownership and error model.

Quick start

# from the repo (until packages are published to npm)
npm install
npm run build

# authenticate (stored 0600 in ~/.vivideo/config.json), or use VIVIDEO_API_KEY
export VIVIDEO_API_KEY="vv_live_..."   # from https://app.vivideo.ai/account/api-keys

# CLI
node packages/cli/dist/index.js account
node packages/cli/dist/index.js create auto --prompt "A 20s product teaser" --wait

# MCP server (stdio) — point an MCP client at this command
node packages/mcp/dist/index.js

Once published: npm i -g @vivideo/cli (gives vivideo), and npx @vivideo/mcp for the server.

Docs

• MCP server — install, configure, connect to clients, tool reference
• CLI — install, auth, command reference, JSON & non-interactive use, exit codes
• Skills — install & use
• Workflows — human & agent, end to end
• Security, rate limits, idempotency, credits & premium, troubleshooting

Authentication & security

• Keys are read from (in order): explicit option → VIVIDEO_API_KEY env → ~/.vivideo/config.json (written owner-only 0600).
• The key is never printed — all output passes through secret redaction; configure stores it without echoing it.
• Keys are account-scoped secrets. Keep them server-side; one per integration so a leak is revocable in isolation.

Not included (honest scope)

• Cancellation — the public API has no cancel endpoint, so no cancel tool/command is offered. Failed/stuck renders are auto-refunded by the API.
• Agent/chat mode — not exposed by the public API, so not in the toolchain.
• Nothing here publishes to npm, provisions DNS, or deploys — those are external steps (see the final section of docs/mcp.md).

Development

npm run build       # build all three packages (tsc -b)
npm run typecheck   # type-check the whole workspace
npm test            # vitest

Requires Node ≥ 18.17.