# CLI Anything > CLI-Anything integration guide for Synthex. Documents how to wrap external tools (FFmpeg, Vercel, Supabase, Stripe, GitHub, Playwright, social platforms) as agent-native CLIs for both internal dev/ops and client-facing automation. Use when adding new CLI-based capabilities to the autopilot or workflow engine, or when wiring a new external tool for agent consumption. - Skill: `cleanexpo/cli-anything` (Agent Skill) - Install (CLI): `npx skillmds@latest add cleanexpo/cli-anything` - Raw SKILL.md: https://api.skillmd.com/api/skills/cleanexpo/cli-anything/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: cleanexpo (https://skillmd.com/u/cleanexpo) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/cleanexpo/cli-anything --- ## Purpose CLI-Anything (HKUDS/CLI-Anything) is a Claude Code plugin that auto-generates agent-native CLI wrappers for any software. Given a target binary, repo, or SDK, it runs a 7-phase pipeline (Analyse → Design → Implement → Test → Document → Publish → SKILL.md) and produces a pip-installable package with structured JSON output and a `SKILL.md` that Claude Code auto-discovers on install. CLIs are the universal agent interface: structured JSON in, structured JSON out, no screen scraping, no natural-language parsing. Synthex uses two layers: 1. **Platform layer (internal)** — wraps dev/ops tools (FFmpeg, Vercel, Supabase, Stripe, GitHub, Playwright) so the autopilot and workflow engine can drive them without human intervention. 2. **Client surface (past the paywall)** — exposes capabilities to Synthex customers as first-class workflow steps: social publishing, GWS automation, competitor research, campaign intelligence, and model switching. --- ## Installation Install the Claude Code plugin once: ```bash /plugin marketplace add HKUDS/CLI-Anything /plugin install cli-anything ``` Generate a CLI for any tool: ```bash /cli-anything:cli-anything ``` After the 7-phase pipeline completes, install the output package: ```bash pip install ./output/ ``` Claude Code auto-discovers the bundled `SKILL.md` immediately — no manual copy to `.claude/skills/` required. The REPL banner confirms the discovery path. --- ## Layer A — Synthex Platform CLIs These CLIs give the Synthex autopilot and workflow engine direct, structured control over the tools the platform already depends on. ### FFmpeg — Video Pipeline > **REAL IMAGES ONLY (founder mandate — `.claude/rules/real-images-only.md`):** ffmpeg > processing now lives on the **Railway media worker**, exposed via `media_*` MCP tools > plus `scripts/` frame extraction. This is the corpus-growth path — extracted frames > from owned videos feed the owned reference library (`public/reference-library/` > manifest; private customer photos via `POST /api/admin/private-refs`) that grounds > `generateImage()`/`generate_image`. **Do not generate a rival ffmpeg CLI wired into > `step-executor.ts`.** If a CLI wrapper is ever justified it must route through the > media worker and land frames in the library ingest path, and any generation step type > must call `generateImage()`/`generateBatch()` — the guard test > `tests/unit/ai/no-direct-image-apis.test.ts` blocks direct provider calls. **Current state:** `lib/video/video-processor.ts` calls `fluent-ffmpeg` sequentially with no queue; errors are parsed from stderr strings. **Generate:** ```bash /cli-anything:cli-anything https://github.com/FFmpeg/FFmpeg ``` **Key commands exposed:** - `transcode --input --codec h264 --output ` - `trim --start --end ` - `add-voiceover --audio --mix ` - `extract-frames --fps 1 --output-dir ` - `watermark --image --position bottom-right` - `concat --manifest ` **Wiring:** Add a `video-process` step type in `lib/workflow/step-executor.ts`. JSON output replaces the stderr-parsing logic in `lib/video/video-orchestrator.ts`. The ElevenLabs voiceover (already in Vercel env) feeds directly into `add-voiceover`. --- ### Vercel — Autonomous Deployments (UNI-1181) **Current state:** Deployments are triggered by `git push` only; no agent control post-push. **Generate:** ```bash /cli-anything:cli-anything https://github.com/vercel/vercel ``` **Key commands exposed:** - `deploy --prod` - `rollback ` - `logs --function --tail` - `env set --environment production` - `env get ` - `domains list` **Wiring:** Add a `deploy` step type in `lib/workflow/`. The hive-mind agent triggers this post-merge. This enables the full autonomous loop: deploy → verify logs → rollback if unhealthy, without a human in the loop. Directly unblocks UNI-1181. --- ### Supabase — DB Agent Control **Current state:** Migrations run manually via `npx prisma db push`; no agent-triggered schema ops. **Generate:** ```bash /cli-anything:cli-anything supabase-cli ``` **Key commands exposed:** - `migrate up` - `migrate down --to ` - `db push --dry-run` - `db diff` - `logs --type api --tail` - `branches list` **Wiring:** Add a `db-migrate` step type. Gate: always run `npx prisma validate` before `supabase migrate up` — this is non-negotiable per the Synthex migration safety rules. The autopilot can validate schema before campaign launch without human intervention. --- ### Stripe — Billing Automation **Current state:** All Stripe interactions are webhook-driven; no agent-triggered billing ops exist. **Generate:** ```bash /cli-anything:cli-anything stripe-cli ``` **Key commands exposed:** - `subscription cancel --customer ` - `subscription pause --customer --duration ` - `invoice list --customer ` - `invoice retry ` - `customer lookup --email ` - `checkout create --price ` **Wiring:** Add a `billing-action` step type. The autopilot can handle dunning flows, trial extensions, and subscription changes as part of a campaign or retention workflow. **Security constraint:** All Stripe CLI calls must be org-scoped. Never allow cross-org customer operations. Validate `org_id` before every billing action — same rule as all Prisma queries. --- ### GitHub — PR & Issue Automation **Current state:** Raw `gh` CLI calls in hooks produce unstructured text output that is not agent-consumable. **Generate:** ```bash /cli-anything:cli-anything gh ``` **Key commands exposed:** - `pr create --title --body --base main` - `pr merge --squash` - `issue comment --body ` - `ci status --repo ` - `release create --notes ` **Wiring:** Replace raw `gh` bash calls in `.claude/hooks/` with structured JSON output via the generated CLI. The hive-mind agent triggers `pr create` post-implementation and then fires a Linear update via the existing MCP hook. --- ### Playwright — Social & Browser Automation **Current state:** Playwright is used for E2E tests only; social platforms use OAuth API SDKs. **Generate:** ```bash /cli-anything:cli-anything https://github.com/microsoft/playwright ``` **Key commands exposed:** - `run --story ` - `screenshot --url --output ` - `fill-form --url --fields ` - `scrape --url --selector ` **Wiring:** Add a `browser-automate` step type as a fallback publisher for platforms without stable APIs (TikTok web, Threads early access). Also used in client onboarding flows requiring browser-based OAuth setup. --- ## Layer B — Content-Delivery CLIs These CLIs are Synthex capabilities used by the autopilot engine to produce and deliver content for Unite-Group's own clients. Synthex is an internal Unite-Group tool — there are no paid tiers or "upgrade" gates; all capabilities are open. They appear as workflow step options in the AI Command Centre. ### Social Platform CLIs **Platforms:** YouTube Studio, LinkedIn, TikTok, Instagram, Pinterest **Generate (example for LinkedIn):** ```bash /cli-anything:cli-anything https://github.com/linkedin/li-api ``` **Key commands per platform:** - `publish --content --platform ` - `analytics pull --period 30d` - `comment reply --post-id --body ` - `hashtag research --niche ` **Wiring:** Each platform CLI becomes an alternative backend in `lib/social/-service.ts` alongside the existing SDK integration. The autopilot falls back to the CLI when the SDK token has expired or rate-limited. **Client benefit:** Agent publishes across all platforms without requiring a new OAuth dance each session. --- ### Google Workspace (GWS) **Platforms:** Gmail, Google Calendar, Google Drive, Google Business Profile (GBP) **Generate:** ```bash /cli-anything:cli-anything google-api-nodejs-client ``` **Key commands exposed:** - `gmail send --to --subject --body ` - `calendar create-event --title --start
--end
` - `drive upload --file --folder-id ` - `gbp post --location-id --text ` - `gbp review-reply --review-id --body ` **Wiring:** Wire into campaign automation workflow: email sequences trigger from campaign milestones, GBP posts schedule via the cron autopilot, Drive delivers reports automatically. **Client benefit:** Fully automated client communication + GBP management from a single workflow, no manual Google login required. --- ### Analytics & Reporting **Generate from existing services:** ```bash /cli-anything:cli-anything lib/ai/insights-generator.ts ``` **Key commands exposed:** - `report generate --period 30d --format pdf` - `anomaly scan --metric engagement --threshold 20%` - `sentiment pull --platform instagram --days 7` - `export --format pdf --destination drive` **Wiring:** Add to the scheduled autopilot cron (`cron/autopilot`) to auto-generate weekly client reports and deliver to their inbox or Drive folder. **Client benefit:** Reports land in the client inbox automatically — no dashboard visit needed. --- ### Competitor Research **Generate from existing tracker:** ```bash /cli-anything:cli-anything lib/social/competitor-tracker.ts ``` **Key commands exposed:** - `competitor track --handle @competitor` - `gap analysis --vs @competitor` - `benchmark --metric follower-growth --period 90d` - `alert set --metric engagement --threshold 10%` **Wiring:** Feed into the autopilot weekly learning loop (`cron/autopilot-learn`). Agents automatically pull competitor deltas and adjust content strategy without human review. **Client benefit:** Agents track competitor moves and adapt the client's content strategy automatically week-over-week. --- ### Campaign Intelligence **Generate from strategy layer:** ```bash /cli-anything:cli-anything lib/autopilot/content-strategy.ts ``` **Key commands exposed:** - `campaign plan --goal awareness --days 30` - `campaign preview --id ` - `performance score --campaign-id ` - `campaign adapt --reason "low engagement"` **Wiring:** Expose as interactive commands in the AI Command Centre (`components/command-centre/`). Agent proposes, client approves once, agent executes the entire campaign without further manual setup. **Client benefit:** Client approves a goal; the agent handles planning, content creation, scheduling, publishing, and performance adaptation end-to-end. --- ### LLM / Model Switching (LLMFit) **Generate from model registry:** ```bash /cli-anything:cli-anything lib/ai/model-registry.ts ``` **Key commands exposed:** - `model list --tier premium` - `model benchmark --task caption --models claude,gpt4` - `model set --tier premium --model claude-opus-4-7` - `cost estimate --campaign-id --model claude-sonnet-4-6` **Wiring:** Integrates into the content pipeline BYOK flow. Clients override the model per campaign from their dashboard. Cost-conscious clients choose cheaper models; power users opt into Claude Opus for quality-critical campaigns. **Client benefit:** Full model transparency and control — clients see cost estimates before committing to a model for a campaign. --- ## Workflow Engine Integration Generated CLIs become first-class step types in `lib/workflow/step-executor.ts`: ```typescript // New step types unlocked by CLI-Anything wrappers: // 'video-process' → cli-anything-ffmpeg // 'deploy' → cli-anything-vercel // 'db-migrate' → cli-anything-supabase // 'billing-action' → cli-anything-stripe // 'browser-automate' → cli-anything-playwright // 'social-publish' → cli-anything- // 'gbp-post' → cli-anything-gws // 'report-generate' → cli-anything-analytics // 'campaign-adapt' → cli-anything-campaign-intel // 'model-switch' → cli-anything-llmfit ``` **Pattern for adding a new step type:** 1. Generate CLI via CLI-Anything: `/cli-anything:cli-anything ` 2. `pip install ./output/` — SKILL.md auto-discovered 3. Add step type handler in `lib/workflow/step-executor.ts` 4. Add step schema in `lib/workflow/types.ts` 5. Write integration test verifying the step runs end-to-end in a workflow --- ## SKILL.md Auto-Discovery Every generated CLI ships a `SKILL.md` inside its Python package. After `pip install`: - Claude Code scans installed packages and discovers embedded `SKILL.md` files automatically - The REPL banner prints the absolute discovery path on first load - No manual copy to `.claude/skills/` is required This means **every new CLI-Anything generation automatically extends the agent's skill set**. The SKILL.md documents the generated CLI's commands, arguments, and JSON output schema so the agent can use it immediately without reading source code. --- ## Priority Order | CLI | Layer | Impact | Effort | Priority | Notes | | -------------- | -------- | ------ | ------ | -------- | ----------------------------- | | FFmpeg | Platform | High | Low | 1 | Already in use — just wrap | | Vercel | Platform | High | Low | 2 | Unblocks UNI-1181 directly | | GitHub | Platform | Medium | Low | 3 | Replaces raw `gh` hook calls | | Supabase | Platform | Medium | Medium | 4 | DB autonomy for migrations | | Stripe | Platform | Medium | Medium | 5 | Billing automation (dunning) | | GWS | Client | High | Medium | 6 | GBP + email campaigns | | Social CLIs | Client | High | High | 7 | Per-platform wrappers (×5) | | Playwright | Both | Medium | Medium | 8 | Browser fallback + onboarding | | Analytics | Client | Medium | Low | 9 | Wraps existing services | | Campaign Intel | Client | High | Medium | 10 | Premium tier feature | --- ## Common Mistakes | Mistake | Correct Approach | | ------------------------------------------------------------------- | ---------------------------------------------------------------- | | Running `npx next build` instead of `npm run build` | Always use `npm run build` — Turbopack vs webpack differ | | Targeting an SDK repo instead of the CLI binary | Target the CLI harness or binary, not the library SDK | | Forgetting `pip install` after generation | SKILL.md is only discovered after `pip install` — always install | | Cross-org Stripe CLI calls | Validate `org_id` before every billing action — never cross-org | | Skipping `prisma validate` before Supabase migrate step | `npx prisma validate` is a hard gate before any `migrate up` | | Wiring a CLI step without a schema entry in `lib/workflow/types.ts` | Add both handler and schema — the type system enforces this | --- ## File Index Key Synthex files this skill references: ``` lib/video/video-processor.ts — FFmpeg integration target lib/video/video-orchestrator.ts — Video workflow (replace stderr parsing) lib/workflow/step-executor.ts — Add new CLI step types here lib/workflow/types.ts — Step type schema definitions lib/autopilot/launch-pipeline.ts — Autopilot launch pipeline (CLI gates) lib/autopilot/performance-learner.ts — Weekly learning loop (CLI input) lib/ai/model-registry.ts — Model registry (LLMFit CLI target) lib/ai/insights-generator.ts — Analytics CLI generation target lib/social/ — Social platform services (9 files) lib/social/competitor-tracker.ts — Competitor research CLI target lib/autopilot/content-strategy.ts — Campaign intelligence CLI target app/api/command-centre/ — AI Command Centre API routes components/command-centre/ — AI Command Centre UI scripts/ — Existing utility scripts to unify ``` > **Reference skill:** This is a read-only architecture guide — it documents existing systems and does not generate creative or code output. No capability uplift block is needed.