Claude Code Skills by deapi-ai

deAPI Skills

Agent skills for deAPI. With them an AI agent can transcribe recordings and video links, generate and edit images, create speech, music and video, read text from images, remove backgrounds, upscale and compute embeddings, and help developers build all of this into their own apps.

The skills follow the agentskills.io open standard, so the same folder works in Claude Code, Codex, Cursor, Hermes Agent, OpenClaw, Goose, OpenCode, GitHub Copilot and other agents that support skills.

Quick start

export DEAPI_API_KEY="sk-dp-..."      # create a key at https://app.deapi.ai/settings/api-keys
npx skills add deapi-ai/skills        # or see "Install" below for your agent

Then ask your agent, for example:

• "Transcribe this YouTube video, it's in English: https://www.youtube.com/watch?v=..."
• "Generate a 1024x768 image of a red fox in studio light"
• "Read this text aloud in my voice, the sample is voice.wav"
• "How do I verify a deAPI webhook signature in Node?"

Every paid job is priced first, for free, and the agent stops and asks you when the price is above your limit (0.25 USD per job by default).

The two skills

Skill	Use it to	How it works
deapi	Get something done now: transcribe or subtitle a file or a YouTube, X, TikTok, Twitch or Kick link; generate, edit and upscale images; remove backgrounds; read text from images (OCR); turn text or a photo into video and describe videos; speech with preset voices, a cloned voice or a designed voice; music; embeddings; better prompts; job status; account and key limits.	A helper script (deapi/scripts/deapi.py, Python 3.9+, standard library only), or the deAPI MCP tools when they are connected.
deapi-api	Build deAPI into an application: the native API v2, async jobs and webhooks, WebSockets, limits and errors, keys and usage, and OpenAI SDKs pointed at https://oai.deapi.ai/v1.	Reference pages and tested code examples (Python, TypeScript, Node, curl). It runs nothing on its own.

Install

1. Get an API key

Sign up at deapi.ai and create a key at app.deapi.ai/settings/api-keys (Settings > API Keys). Keys look like sk-dp-<id>-<secret>; use them as they are.

For an agent, create a dedicated key with the API role and a daily spend limit. deAPI enforces that limit itself, whatever the agent does.

2. Make the key available

The skills read the key from the DEAPI_API_KEY environment variable. Never paste the key into the chat.

export DEAPI_API_KEY="sk-dp-..."      # add it to your shell profile

3. Install the skills

Any agent, via skills.sh:

npx skills add deapi-ai/skills

Claude Code (manual):

git clone https://github.com/deapi-ai/skills.git
mkdir -p ~/.claude/skills && cp -r skills/deapi skills/deapi-api ~/.claude/skills/

To let Claude Code run the helper without asking each time, allow only that script in settings.json, for example "allow": ["Bash(python /path/to/deapi/scripts/deapi.py *)"] with the path where you installed the skill. The skill tells the agent to run one command per call, so the rule matches.

Hermes Agent:

hermes skills install deapi-ai/skills/deapi

Hermes does not ask for the key when it loads the skill. Add it to the Hermes .env file (~/.hermes/.env by default; hermes config env-path prints the path), which Hermes passes to the terminal where the script runs:

DEAPI_API_KEY=sk-dp-...

If your deAPI account has access to chat models, Hermes can also use one as its main model through the OpenAI-compatible API (hermes model, "Custom endpoint", or in config.yaml):

model:
  provider: custom
  base_url: https://oai.deapi.ai/v1
  key_env: DEAPI_API_KEY
  default: <a chat model ID from GET https://oai.deapi.ai/v1/models>

A text-only chat model cannot handle Hermes' image tasks (vision); point auxiliary.vision at another provider if you need them.

To make deAPI the engine behind Hermes' built-in image_generate and video_generate tools, install the deAPI plugin (hermes plugins install deapi-ai/hermes-plugin-deapi, hermes plugins enable deapi, then pick deAPI in hermes tools). For the deAPI MCP tools, add the server to config.yaml; Hermes fills in the key from its .env:

mcp_servers:
  deapi:
    url: https://mcp.deapi.ai/mcp
    headers:
      Authorization: "Bearer ${DEAPI_API_KEY}"

Codex, Cursor, Goose, OpenCode, VS Code Copilot (shared ~/.agents/skills convention):

git clone https://github.com/deapi-ai/skills.git
mkdir -p ~/.agents/skills && cp -r skills/deapi skills/deapi-api ~/.agents/skills/

OpenClaw: copy the deapi folder to the agent's skills folder. The skill declares that it needs DEAPI_API_KEY and one of python3, python or py.

Requirements: the deapi skill needs Python 3.9 or newer (the version macOS ships) and network access to api.deapi.ai. On Windows, python or py works too.

Use

Ask your agent in plain language. You can also run the helper script yourself:

python3 deapi/scripts/deapi.py doctor                         # check Python, the key and limits
python3 deapi/scripts/deapi.py transcribe "https://www.youtube.com/watch?v=..." --lang en
python3 deapi/scripts/deapi.py image --prompt "a red fox, studio light" --quote
python3 deapi/scripts/deapi.py transcribe --duration 7200 --quote   # price of 2 hours, no file needed
python3 deapi/scripts/deapi.py models --type image
python3 deapi/scripts/deapi.py account

Command	What it does	Cost
transcribe (alias stt)	Transcript of a file or a YouTube, X, TikTok, Twitch or Kick link; language, timestamps, speakers, --subtitles srt|vtt	paid
image, edit	Image from text; edit one or more images with a prompt	paid
ocr, rmbg, upscale	Text from an image; background removal; image upscaling	paid
video, animate	Video from text; animate an image (first and last frame, or references)	paid
audio-video, video-upscale	Video that follows an audio track; video upscaling	paid
describe	Scenes and events of a video file or link	paid
tts, music	Speech (preset voice, cloned voice, designed voice); music with or without lyrics	paid
embed, boost	Text embeddings; a better generation prompt	paid
models, status, account (alias balance), doctor	Model catalog, job status and download, key limits and usage, setup check	free

Every paid command takes --quote (free price, nothing submitted), --max-cost, --model and --output. deapi.py <command> --help lists all flags with examples; the full reference, including exit codes, is in deapi/references/commands.md.

Cost control

• Quote first. Before any paid job the script asks deAPI for the price (free) and prints it together with max_cost, the limit in force.
• Limit per job. Above the limit (0.25 USD by default) the script stops with exit code 5 and the agent asks you. Only you raise it, for one job, with --max-cost set to the approved price or with DEAPI_MAX_COST. The skill tells the agent never to raise it on its own.
• Limit per key. A daily spend limit on the key (set in the deAPI dashboard) caps everything the agent can spend, independent of the skill.
• No double charges by accident. A submitted job is retried only when deAPI answers "rate limited" (HTTP 429, nothing was charged) or when it fails and deAPI refunds it (once, after a new quote).

Tested platforms

Platform	Result	Tested on
Claude Code	The skill loads for media tasks and deapi-api for app questions; transcription, image, balance, price-only questions and a job above the limit (the agent stopped and asked) all behaved as described.	Claude Code 2.1, November 2026
Hermes Agent	Balance, image, price-only question and speech with a deAPI chat model as Hermes' main model; the Hermes skill scanner rates both skills safe.	Hermes Agent in Docker, November 2026
OpenClaw	The skill and its requirements are read correctly by the OpenClaw loader and by ClawHub, also when only python is installed.	Loader check, November 2026
Codex, Cursor, Goose, OpenCode, Copilot, Windsurf	Expected to work through the same folder and script; not tested yet.	—
claude.ai, ChatGPT and other chat apps	Skills that run scripts are not supported there; use the deAPI MCP connector instead.	—

Where your data goes

• To deAPI, over HTTPS: the prompts and parameters of a job and the files you point the agent at go to api.deapi.ai. Results come back from deAPI result links, which the script downloads without your key. With the MCP tools the same data goes to mcp.deapi.ai; apps built with the OpenAI SDKs talk to oai.deapi.ai.
• Your key is read from DEAPI_API_KEY and sent only to *.deapi.ai hosts; the script refuses any other base URL and never lets the key follow a redirect. It is never written to disk or printed.
• Your files: only the files you point the agent at are sent. Hidden files and folders such as .ssh, .aws or .env are refused, also for text inputs and output paths, unless you name the file and the agent adds --allow-hidden-path.
• Where jobs run: deAPI runs models on a decentralized GPU network. Do not send recordings, documents or images that hold sensitive personal data or trade secrets; see Architecture and security.
• Nothing else: no telemetry or analytics. Each request carries the header User-Agent: deapi-skill/<version>. Results are saved in the folder you choose (--out-dir).

Other ways to use deAPI

• MCP server: https://mcp.deapi.ai/mcp exposes the same jobs as MCP tools (deapi-ai/mcp-server-deapi). Use it in clients without a terminal, such as claude.ai.
• OpenAI-compatible API: https://oai.deapi.ai/v1 works with OpenAI SDKs for images, speech, transcriptions, embeddings and videos.
• Native REST API: docs.deapi.ai (OpenAPI: openapi.json).
• Python SDK: deapi-ai/deapi-python-sdk wraps every endpoint, price quotes, polling, webhook verification and WebSockets. It needs Python 3.11+ and installs from GitHub: pip install "git+https://github.com/deapi-ai/deapi-python-sdk.git@v0.5.2" (pip install deapi is an unrelated package).

Updates and versions

• Each skill carries its version in metadata.version of its SKILL.md; releases are tagged vX.Y.Z with notes in CHANGELOG.md and on the Releases page.
• Installed with npx skills add: update with npx skills update.
• Installed by copying the folder: pull the repository again and copy the folder over the old one.
• If the script exits with code 13 ("endpoint gone"), the skill is outdated: update it.

Coming from deapi-ai/claude-code-skills?

This repository is the former deapi-ai/claude-code-skills (renamed; old links redirect here). The old Claude Code commands used the deprecated deAPI API v1. They are kept for reference at the v1-final tag and are no longer maintained.

To switch:

1. Remove the old copy: ~/.claude/skills/deapi and any deAPI command files you copied (~/.claude/commands/, .cursor/commands/, .windsurf/workflows/).
2. Install the current skills (see Install).
3. Keep your DEAPI_API_KEY; it works with the new skills.

The slash commands are gone. Ask your agent in plain language and the deapi skill picks the right command:

Old command	Now
/transcribe	transcribe
/generate-image	image
/generate-audio	tts
/generate-video	video, animate
/ocr	ocr
/remove-bg	rmbg
/upscale	upscale
/transform-image	edit
/embed	embed
/deapi-setup (webhooks, WebSockets)	the deapi-api skill
/deapi-balance	account
deapi-audio skill	tts, transcribe

Repository layout

deapi/                 the "do it now" skill: SKILL.md, scripts/deapi.py, references/
deapi-api/             the "build it into an app" skill: SKILL.md, references/, assets/
tests/                 offline unit tests, synthetic fixtures, Dockerfile
tools/                 repository checks (frontmatter, OpenAPI, MCP names, skill safety, leak scan)
skills.sh.json         skill groupings for skills.sh and the Hermes Skills Hub

Contributing: see AGENTS.md.

License

MIT