Use inference.club from AI agents
Пока недоступно на языке Русский — показан английский.

Use inference.club from AI agents

There are two ready-made integrations. Both authenticate with your regular API key (Dashboard → Settings → Token) and both are capacity-aware: they send "async": "auto", so a busy GPU queues the request instead of getting overloaded, and they wait for the job and hand back media URLs.

Skill + CLIRemote MCP server
Best forAgents with a shell (Claude Code, Hermes, Codex, OpenClaw…)Clients without a shell (Claude Desktop, claude.ai, Cursor) or MCP-first agents
Saves files locallyYes (-o out.png)No — returns URLs
InstallCopy a folderAdd one URL + header
Lives inintegrations/skills/inference-club/ in the repohttps://api.inference.club/mcp

The skill is a SKILL.md plus ic.py, a dependency-free Python CLI that wraps the whole API. Each command prints one JSON object, so agents parse results reliably.

git clone https://github.com/inference-club/inference.club && cd inference.club
integrations/skills/install.sh                    # Claude Code → ~/.claude/skills/
integrations/skills/install.sh --hermes           # Hermes default profile
integrations/skills/install.sh --hermes minimal   # a specific Hermes profile

The installer also links the CLI as ~/.local/bin/inference-club. Then give the agent your key:

  • Claude Code: export INFERENCE_CLUB_API_KEY=... in the shell you launch claude from.
  • Hermes: add INFERENCE_CLUB_API_KEY=... to ~/.hermes/.env (or the profile's .env). The skill declares it under required_environment_variables, so Hermes passes it to the terminal tool.

Self-hosting or local dev? Also set INFERENCE_CLUB_BASE_URL (e.g. http://localhost:8101).

Now just ask: "generate an image of a lighthouse at sunset with inference club and save it as lighthouse.png". You can also use the CLI yourself:

inference-club doctor                     # checks the key (masked) + connectivity
inference-club models --type image
inference-club image "a watercolor fox" -o fox.png
inference-club video "slow dolly shot through a misty forest" -o forest.mp4
inference-club music "lofi hip hop, mellow piano" --duration 60 -o song.mp3
inference-club speech "Welcome to the show." -o hello.wav
inference-club chat "Explain diffusion models in two sentences"
inference-club job 123 --wait -o out/     # resume a queued job
inference-club capacity                   # live GPU load

Option 2 — the remote MCP server

POST https://api.inference.club/mcp speaks the Model Context Protocol (Streamable HTTP, stateless JSON responses) with Bearer auth.

Claude Code

claude mcp add --transport http inference-club https://api.inference.club/mcp \
  --header "Authorization: Bearer $INFERENCE_CLUB_API_KEY"

Hermes (config.yaml)

mcp_servers:
  inference-club:
    url: https://api.inference.club/mcp
    headers:
      Authorization: "Bearer ${INFERENCE_CLUB_API_KEY}"

Any other client: add a remote/HTTP MCP server with that URL and an Authorization: Bearer <key> header.

Tools: list_models, get_capacity, generate_image, generate_video, generate_music, generate_speech, chat, scrape_url, get_job, list_jobs, cancel_job, list_workflow_templates, run_workflow, get_workflow_run.

A generation tool waits up to ~45 seconds. If the job hasn't finished by then (typical for video), it returns status: "QUEUED" with a job_id; the agent calls get_job later. Results include the media URLs as text, structuredContent, and MCP resource_link items.

How agents should behave

These rules are built into both integrations, and are worth following if you call the raw API yourself:

  • Discover models with GET /v1/models and choose by service_type.
  • Send "async": "auto": the request runs now if a slot is free, otherwise it's queued (202, queued_reason: at_capacity | no_provider). Poll poll_url with backoff.
  • Send "delivery": "url" to audio/video endpoints to get JSON with URLs instead of bytes.
  • 429 at_capacity means the device is busy and the request wasn't queued. Wait for Retry-After, or resend with "async": "auto".
  • Don't resubmit a queued request; poll its job instead.
  • GET /v1/capacity shows live load if you want to plan ahead.

The machine-readable spec is at /openapi.json, and a compact primer for LLMs is at /llms.txt.