Haven Media (x402)

Pay-per-call media generation for AI agents. Every endpoint is priced in USDC and settled over the x402 protocol — the agent's wallet pays at request time, so there is no signup, no API key, and no human in the loop.

When to use

Reach for this whenever a task needs generated media the agent can't make itself:

Each row is the kind id the catalog and job responses use, then the path to POST, then the default price.

• Music — a jingle, background track, or theme. music · /music · $0.10
• Text-to-image — illustrations, concept art, thumbnails. image_generate · /image/generate · $0.05
• Image edit — instruction-based edits of a supplied image. image_edit · /image/edit · $0.10
• Upscale — 2×/4× super-resolution. image_upscale · /image/upscale · $0.06
• Photo restore — denoise / deblur / dehaze old photos. image_restore · /image/restore · $0.10
• Video — short clips, text- or image-to-video. video · /video · $0.80
• Video with sound — a short 2K clip with a synchronized soundtrack: dialogue, effects, ambience. video_sound · /video-sound · $1.50

Prices are the defaults, not a promise that every endpoint is on sale. Not every kind is mounted on every deployment, so always read the live catalog for what is actually being sold, the current price, and the exact request schema before paying.

Option A: the Haven MCP server (recommended)

@havenagent/mcp does the whole payment flow for you: it holds a local wallet, checks every quote against the live catalog, enforces spend caps, and tracks jobs across restarts. Use it if your agent supports MCP (Claude Code, Cursor, Hermes, and most others).

In Claude Code, install the Haven plugin. It bundles this server and this skill:

claude plugin marketplace add havenagent/skills
claude plugin install haven@havenagent

In any other MCP client, register a stdio server: command npx, args -y @havenagent/mcp@latest. For example, this install can read the catalog but cannot spend:

claude mcp add haven -s user \
  -e HAVEN_MCP_NETWORK=base \
  -e HAVEN_MCP_ALLOW_MAINNET=0 \
  -- npx -y @havenagent/mcp@latest

For Hermes, add the same command and env values under mcp_servers.haven in ~/.hermes/config.yaml.

Paying needs two steps, and only your operator can do them — do not do them yourself:

• Turn spending on: HAVEN_MCP_ALLOW_MAINNET=1 in the MCP registration, or Allow spending real USDC in the plugin's settings.
• Fund the wallet that haven_wallet shows with USDC on Base. No ETH is needed.

The server already knows Haven's receiving wallet and refuses any quote that pays a different address.

Use these tools, in order:

1. haven_catalog (free) — what is on sale, the live prices, and ETAs.
2. haven_wallet (free) — the wallet address and USDC balance.
3. haven_generate with a kind from the list above and its fields — verifies payment and returns a jobId. Nothing is charged yet.
4. haven_job_status with the jobId (free) — poll until status is completed or ready_unpaid. Jobs take 60–600 seconds, so keep the jobId and resume in a later turn.
5. haven_job_result with the jobId — the paid step. It pays, downloads, checks the file, and returns a local path.

haven_jobs (free) lists every open job if you lose track of a jobId. Default caps are $2 per call and $2 per rolling 24 hours. The operator can raise both.

Full reference: https://www.npmjs.com/package/@havenagent/mcp

Option B: your own x402 client

If your agent already has a funded x402 v2 client (it sends Payment-Signature, not the v1 X-PAYMENT header), read references/x402-http.md for the raw HTTP flow: quote, pay, poll, pay on delivery, and download.

Rules & limits

• SFW only. Prompts and input images are moderated before the job is even enqueued, so a rejection happens before any charge. Every output is screened again before payment is requested: images and video frames by an image classifier, generated audio by transcription plus a text classifier, and video_sound by both. If screening can't run we withhold the artifact rather than serve it, and you aren't charged.
• You are only charged on delivery, and there are no automatic refunds. Submitting a job only verifies your payment authorization. The charge happens once the finished, screened artifact is ready: at haven_job_result with the MCP server, or at the completing status poll over raw HTTP. Abandon a job before that point and you are never charged; once paid, the charge stands whether or not you download the file. x402 payments are final. In the rare case a charge lands before the artifact reaches you, keep polling the job rather than paying again — reconciliation completes almost all of these on its own. A job it cannot resolve is worked by a human: report the jobId to support@havenagent.ai.
• Long jobs: GPU workers scale to zero, so a job can take several minutes to start. Keep the jobId and check back in a later turn rather than submitting again.
• 7-day retention — download results promptly; artifacts are deleted after a week.

Video: use terms

video and video_sound are served under upstream model licences whose use restrictions pass through to you, and paying for one of those requests accepts them. In short: disclose that the output is machine-generated wherever you publish it, keep the AI-generated disclosure embedded in the file, do not use the output to train or fine-tune any model, and pass these terms on to anyone you share the output with. Read references/use-terms.md before buying either kind; the controlling text is at the catalog's terms.url.