Skip to main content

Primex Skill

Give your agent a task and the Primex skill. It finds a suitable model, gets its current inputs and rates, then runs the task and delivers the result. Keep your existing agent and reasoning provider while using Primex for images, speech, transcription, documents or translation.

Service use needs HTTPS tools, private credential settings and file access when the task involves files. Python, a downloaded package and a local server are optional.

Read the skill · Download version 0.1.0 · Archive checksum

Try a task​

Start with a request such as:

Read https://docs.primexos.com/SKILL.md and use Primex to find a narration model. Show me its voices and rates before generating audio.

Other examples:

  • “Find a Primex speech model and show me its voices and price.”
  • “Generate one image through Primex from this prompt and save it here.”
  • “Transcribe this recording with Primex, then translate it into German.”
  • “Extract the fields from this invoice using Primex.”
  • “Configure Codex to use this Primex model in this project.”

Search and inspection do not run inference. Generating or processing content spends Primex credits; rates are not a final quote. The agent asks for missing choices that affect your result or cost.

How it works​

  1. Find a capability. Public task search returns a small set of model and operation references, selection information, rates and the next HTTP request.
  2. Check its requirements. Non-billable Gateway inspection returns the current input schema, model limits, voices where relevant, upload steps and output instructions. It accepts your privately configured key even at zero balance. For missing ElevenLabs voice attributes, an optional non-billable lookup helps choose a voice accepted by the inspected model.
  3. Run and deliver. The agent fills the returned request, submits the task once and returns the actual text, structured data, image or audio.

Without a key, anonymous previews show the public requirements. Runtime inspection is still needed before execution. Inspection does not reserve funding, price or capacity; the execution endpoint checks them again.

The HTTP workflow explains filters, pagination and request handling. There is no separate skill server to run.

Connect your account​

Create a dedicated key in API Keys and configure PRIMEX_API_KEY through your host's private secret or environment settings. Set an expiry and monthly cap if appropriate; see the cap's in-flight behavior. Never paste the key into ordinary chat.

The key goes only to the Primex Gateway. File uploads use their returned storage headers. Manage funding and usage in the App; your agent subscription and Primex API usage have separate billing.

In Codex and Claude Code, the key must be available to the HTTP executor. Existing shell environment policies can exclude it even when the CLI starts successfully. Check availability without printing the key and preserve your existing policies; see the connection guide.

Reading the skill URL makes its instructions available for the current task. It does not install a persistent skill or configure your account.

Keep the skill available​

For future tasks, download and inspect the archive, then install its primex directory using your agent's skill mechanism. Preserve existing skills and configuration.

HostInstallation
Claude CodePlace primex under .claude/skills/ for one project or ~/.claude/skills/ for your user. Invoke /primex.
CodexPlace primex under .agents/skills/ in your project. Invoke $primex; see the official installation locations.
Other skill hostsUse their documented package import. HTTP tools, file access and private secret settings vary by host.

The package has a manifest.json listing hashes for its contents. The download checksum checks archive integrity; it is not an independent signature. When updating, review the new version and preserve local configuration.

Supported tasks​

The discovery and inspection contracts cover single-image generation, MP3 speech, transcription, OCR, structured document analysis and text translation. Model availability and rates come from current Primex data. Other capabilities link to their API guides; a listed model does not make every upstream feature available.

Inspection supplies the chosen operation's upload size and MIME limits. The agent keeps the operation ID and UTC submission time, then the request ID and returned cost. If delivery fails after a response arrives, recover that response before submitting again. See task recipes and files and recovery.

For client setup, the skill follows the relevant guide and preserves other providers and approvals. Checking settings and successfully running a client task are separate steps.

Optional local helper​

The package includes commands for discovery, inspection, file saving, uploads and client configuration checks. Only this helper needs Python 3.11 or newer; it needs no additional Python packages. It consumes the same server contracts as the HTTP workflow and saves private receipts. Its local upload bound is 25 MiB; server limits can be lower. See the helper reference.

Need help? Report a setup issue with your host/version, operation and safe request ID. Keep keys and private inputs out of ordinary support messages.