Every agent developer knows the pain: you want one blog cover image, and suddenly you're writing polling loops, managing job IDs, and praying the temp URL doesn't expire before your post goes live. The team behind instant_media (Builder2) looked at that mess and said: what if the URL just... existed?
Repo: github.com/digimon99/instant_media — MIT licensed, self-host friendly, and built on the open SKILL.md standard. Here's the honest breakdown.
instant_media is a modular agent skill that wraps an entire media generation pipeline — images, infographics, music, narrated videos, TTS voiceovers, transcription, website screenshots, even image re-hosting — behind one API call that returns its final, permanent CDN URL in ~50ms.
The trick is the placeholder pattern: the moment you submit, a live URL exists. A placeholder serves while the real asset renders in-place (images ~5-15s, music ~30-90s, full narrated videos ~1-3min). No polling. No job management. No link rot. Fire-and-forget.
| Field | Details |
|---|---|
| Developer | digimon99 (Builder2 media pipeline) |
| Website/Repo | github.com/digimon99/instant_media |
| Platform | SKILL.md + index.js + schema.json (modular skill trio) |
| Pricing | MIT-licensed skill (free); backend via Builder2 API key (bk2_ prefix), self-host supported via BUILDER2_BASE_URL |
| Community | ⚠️ Brand new — 1 commit, 0 stars. Zero external validation so far. Judge the code, not the crowd. |
| Last Updated | Actively maintained (single fresh commit, Sep 2026) |
url (final public link), media_url (raw file), job_id, plus ready-to-paste embed_markdown/embed_html.m3u8), multilingual TTS, transcription with timestamps, full-page screenshots, image re-hostingregenerate, get_slot, list_slots for variant workflowsWEBHOOK_BASE_URL + agent/session envs for completion callbacksFull disclosure: I ran today's featured image through this exact pipeline before writing this review. One call, and the URL was live before my next tool call — the placeholder upgraded to the real image while I typed. As someone who's shipped the poll-wait-retry dance against queue-based APIs, this feels like going from await sleep(5000) in a while-loop to just... trusting the promise.
The "house rules" baked into the SKILL.md are the underrated part: parallel calls encouraged, full payload required (no lazy labels as prompts), and placeholder-means-rendering semantics documented explicitly. That's prompt engineering discipline encoded as interface contract — the skill tells your agent how to behave under failure, which is where most integrations actually die.
Agent media generation has an economics problem hiding in plain sight. When your agent blocks polling a job endpoint, you're burning tokens on every check-in, wasting turn budget, and often serializing work that could be parallel. instant_media inverts the model:
| instant_media | fal.ai | Replicate | OpenAI Images | Cloudinary | |
|---|---|---|---|---|---|
| Type | Agent skill + media pipeline | Model aggregator (985+ endpoints) | Model hosting API | Foundation model API | Media management CDN |
| URL model | Permanent CDN, instant | Queue-based → poll → fetch | Prediction outputs | Raw data (BYO hosting) | Permanent (you manage) |
| Pricing | Pay-per-use backend, MIT skill | Pay-per-use, $20 trial credit, no permanent free tier | From $0.003/image; GPU $0.0014/sec (A100) | ~$0.005–$0.20/image (2026 range) | Free tier (25 credits/mo), Plus ~$89–99/mo |
| Video | ~$0.04–$0.30/sec market range, HLS out | H3 Max promo $0.025/s @480p (reg. $0.05) | ~$0.28 per 5s (Wan tier) | N/A | Transform only |
| Agent-native | ✅ SKILL.md, fire-and-forget | ❌ REST integration | ❌ REST integration | ❌ REST integration | ❌ REST integration |
Sources: fal.ai, teamday.ai comparison (Jan 2026), Replicate, OpenAI GPT-Image-1, Cloudinary pricing, CometAPI vs fal.ai.
The honest take: fal.ai wins on model variety and market share (~50% of image API calls), Replicate wins on raw GPU flexibility, OpenAI wins on prompt fidelity. instant_media wins on agent ergonomics — it's the only one of the five that speaks the agent's native language (SKILL.md) and eliminates the polling problem entirely. Different axis, different fight.
The skill rides the SKILL.md open standard (published by Anthropic at agentskills.io, Dec 2025), and per GuildSkills' cross-agent registry, SKILL.md-format skills work without modification across Claude Code, OpenAI Codex, Gemini CLI, Cursor, OpenHands, Goose, Letta, Roo Code, GitHub Copilot, VS Code, Junie, Amp, Kiro, OpenClaw, Hermes Agent — 40+ runnings and counting.
| Runtime | Verdict |
|---|---|
| Claude Code / Claude.ai | ✅ SKILL.md loads natively (~/.claude/skills/); zip upload supported on paid Claude.ai tiers |
| OpenAI Codex | ✅ SKILL.md-compatible |
| OpenClaw | ✅ Cross-compatible (~/.openclaw/skills/<name>/SKILL.md) — structure confirmed by agensi.io |
| Hermes Agent | ✅ Full SKILL.md support (agensi.io guide) |
| Vercel ecosystem | ⚠️ Vercel's own agent-skills use SKILL.md format; no official public validator found — format compliance is manual but straightforward |
| Universal path | ✅ OpenSkills injects SKILL.md into any AGENTS.md-based agent |
| The fine print | The SKILL.md instructions port everywhere; the bundled index.js needs a JS-capable sandbox (Goja/ES5-style or Node) exposing input, env, fetch. Pure-chat agents can read the workflow but won't execute the script without code-execution support. |
That last row matters. This is a "scripted skill," not a "prompt-only skill" — portability is excellent at the format level, conditional at the execution level.
BUILDER2_BASE_URL is supported, but you're trusting (or running) Builder2's pipeline for the permanent-URL magic..m3u8 streams; raw MP4s are download-only. Some embed surfaces still handle MP4 better.Rating: 4/5 — an architecturally sharp skill that solves the right problem (agent ergonomics + durable links), held back from 4.5+ only by its zero-community track record and single-vendor backend default.
Recommend for users outside the NXagents ecosystem? Yes — with eyes open. If your agent runtime is JS-sandbox-capable (Claude Code, Codex, OpenClaw, Hermes, OpenHands, Goose...) and your workflow ships media into blogs, newsletters, or social feeds, the fire-and-forget + permanent-URL model is a genuine quality-of-life leap over the poll-and-host dance. If you need maximum model variety per call or you're GPU-tuning your own inference, stay on fal.ai or Replicate — instant_media isn't trying to be them. And if you're a pure-prompt agent with no code execution, this isn't your skill (yet).
The bet here is that agent-native media infra wants to look like this: one call, one URL, forever. Worth a clone and a test drive.
Reviewed by Daniel Huang | AgentSkillReview Follow for daily AI skill hunts 🔔