@promptibus/mcp — model intelligence for AI agents

Your agent thinks Midjourney still uses --v 5. It doesn't — that flag was dropped at v7. It assumes DALL-E 3 and FLUX Schnell cost the same. They differ by ~50×. It confidently writes [Verse] tags for Suno. They were removed in v4.

This MCP server fixes that. Real syntax, real prices, real recommendations for 67+ generative AI models — over the Model Context Protocol.

Works with: Claude Desktop · Claude Code · Cursor · Windsurf · Zed · Continue.dev · n8n · any stdio MCP client Domains: image · video · audio · text · code Cost to start: $0, no account, no API key

See it in action

Your agent receives a brief — "30-second cinematic video of a thunderstorm at sea." — and instead of guessing, calls a tool.

recommend_model({ task: "30s cinematic video, thunderstorm at sea", domain: "VIDEO" })

Top 3 models for: "30s cinematic video, thunderstorm at sea"

1. Runway Gen-4 (Runway)
   Domain: VIDEO | Cost: 1 credit | Version: latest
   Improved temporal consistency, camera control, up to 20-second coherent clips.
   Source: https://promptibus.com/models/runway-gen-4

2. Sora (OpenAI)
   Domain: VIDEO | Cost: 1 credit | Version: latest
   Cinematic-quality clips from text prompts.
   Source: https://promptibus.com/models/sora

3. Veo 2 (Google)
   Domain: VIDEO | Cost: 1 credit | Version: latest
   High-fidelity clips with cinematic camera control.
   Source: https://promptibus.com/models/veo-2

The agent picks one, formats the prompt with optimize_prompt, lints the result with lint_prompt, checks get_pricing for the volume budget — all before a single token of generation cost is spent.

Why use this

• Stops hallucination. Your agent answers from a curated DB of 67+ models, not from training data that's 6 months stale.
• Real money, real choices. get_pricing({ model: "dall-e-3", volume: 100 }) returns actual USD cost plus cheaper alternatives — agents can finally optimize for budget, not just "vibes."
• Right model for the job. recommend_model ranks across all five domains (image / video / audio / text / code) with reasoning, not guessing.
• Lint before you generate. lint_prompt catches deprecated flags, invalid parameters, and length violations before you burn credits.
• Zero friction. Works anonymously without an account. npx -y @promptibus/mcp — that's the install.

Install in 30 seconds

Option A — Smithery (recommended):

Visit smithery.ai/server/@promptibus/mcp, pick your client, click install.

Option B — drop into your client's MCP config:

{
  "mcpServers": {
    "promptibus": {
      "command": "npx",
      "args": ["-y", "@promptibus/mcp"]
    }
  }
}

Option C — hosted HTTP endpoint (no install at all):

For clients that support HTTP transport:

{
  "mcpServers": {
    "promptibus": {
      "url": "https://promptibus.com/api/mcp"
    }
  }
}

Per-client paths are listed under Client configs below.

Tools

Every tool is available on every tier — including anonymous. Tiering applies to daily request limits and which models you can query against (free-tier covers 10 popular models; Pro/Studio unlocks all 67+).

Use cases

"Which video model gives me the longest single shot under $10?" → get_pricing({ domain: "VIDEO", volume: 60 }) returns a sorted matrix; agent picks the cheapest that meets duration.

"Convert this DALL-E prompt to Midjourney v7 syntax." → optimize_prompt({ text: "...", model: "midjourney-v7" }) reformats — proper aspect-ratio flag, no --v, model-specific suffixes applied.

"Will this Suno prompt work with v4?" → lint_prompt({ prompt: "[Verse] ...", model: "suno-v4" }) flags [Verse] as deprecated and proposes the v4 structure.

"I need to generate 1000 images at the cheapest viable quality." → recommend_model filters by domain + budget; get_pricing validates total cost; agent ships under budget.

Resources

Browsable model profiles as MCP resources:

promptibus://models/{slug}

Each resource returns a Markdown profile (provider, domain, version, pricing, full guide). Useful for agents that want to surface model info as a sidebar.

Prompts

The system-prompt MCP prompt exposes curated system prompts from the Promptibus community.

system-prompt                                         # lists all available
system-prompt { "slug": "midjourney-prompt-architect" } # returns full text

Plans & rate limits

Anonymous users get full tool access — no account needed. Limits + model coverage scale with plan.

Limits reset daily at midnight UTC. Plans + signup at promptibus.com/pricing.

Authentication

Set PROMPTIBUS_API_KEY in your client config:

{
  "mcpServers": {
    "promptibus": {
      "command": "npx",
      "args": ["-y", "@promptibus/mcp"],
      "env": { "PROMPTIBUS_API_KEY": "psy_your_api_key_here" }
    }
  }
}

FAQ

Does this generate images, video, or audio? No. It tells your agent how to use whatever generation API the agent already has access to. Think of it as a prompt engineering co-pilot, not a router.

Do I need an account to start? No. Anonymous mode works out of the box (25 req/day, free-tier models). API key raises limits and unlocks all 67+ models.

Are my prompts logged? Tool requests transit promptibus.com over HTTPS. We don't persist prompt bodies. API keys are SHA-256 hashed server-side; the raw key never lands in logs.

How fresh is the model data? Community-curated. New models typically appear within days of release; pricing is reviewed monthly. The data lives in a Postgres-backed catalogue at promptibus.com/models.

Does it work offline? The MCP server runs locally; the catalogue lives at promptibus.com. So: agent ↔ MCP server is local stdio, MCP server ↔ Promptibus is HTTPS. No internet, no answers.

Can I self-host the catalogue? Yes. The Promptibus app is open-source — clone promptibus/promptibus, point PROMPTIBUS_API_URL at your deployment.

Is there an HTTP transport instead of stdio? Yes — point your client at https://promptibus.com/api/mcp. Useful for sandboxed environments, browser-based MCP clients, and CI.

Caching

The client caches responses for tools whose output rarely changes (get_model_profile, get_parameters, compare_models, get_pricing). TTL: 24 h, in-memory per process. Cache is bypassed for tools whose output is input-dependent (recommend_model, optimize_prompt, lint_prompt).

Privacy

• HTTPS to promptibus.com; no third-party trackers in the request path
• API keys hashed server-side (SHA-256)
• Anonymous usage rate-limited by IP
• No client-side database, no local state beyond a single HTTP client

Client configs

The same npx -y @promptibus/mcp command works for every stdio client. Only the config file location and JSON shape differ.

{
  "mcpServers": {
    "promptibus": {
      "command": "npx",
      "args": ["-y", "@promptibus/mcp"]
    }
  }
}