Building with AI coding tools

LLMs are great at writing Hyperclip integrations once they can see the docs. This page covers three complementary paths, from least to most effort:

One-click bundle — copy every doc page as a single prompt-ready blob, then paste into any chat. Best for one-off coding sessions.
Hosted MCP server — connect any AI client to a live, queryable view of this documentation in one URL paste. Best for ongoing work.
Manual context priming — copy individual snippets directly into your chat / Cursor / Copilot session. Best for targeted questions.

Option 1 — One-click: copy the whole docs to your LLM

Bundles every page in this documentation site into a single markdown blob (/llms-full.txt) and writes it to your clipboard. Paste into Claude, Cursor, Copilot, ChatGPT, or any LLM chat and ask it to build with the Hyperclip API. The bundle includes the API reference, SDK reference, node catalog, model catalog, error codes, and pricing — i.e. everything an LLM that has never heard of Hyperclip would need.

Open the bundle in a new tab

The bundle regenerates on every docs deploy, so it always reflects the current docs. There’s also a slimmer table-of-contents version at /llms.txt that most LLM-aware tools recognize automatically.

Suggested first prompt

After pasting the bundle into a fresh chat, kick the model off with something like:

You now have the complete Hyperclip Docs in context above. Hyperclip is
a visual node-based platform for generating short-form video; the API
lets me run flows programmatically. I want to <DESCRIBE YOUR GOAL>.
Using the docs, write a complete <TypeScript / Python / curl> solution.
Cite which doc pages back each design choice.

The “cite which pages” framing forces the model to ground its output in the docs rather than hallucinating endpoints. If it can’t find something in context, it’ll tell you, and you can supplement from the MCP server below.

The bundled file is ~150–200k characters. That fits comfortably in any modern frontier model’s context window (Claude, GPT-4o, Gemini 1.5+). For smaller-context models, use the MCP server or selective priming below.

The other two paths below are more efficient once you’re past the first integration.

Option 2 — Hosted MCP server

There’s a hosted Model Context Protocol server for these docs. Any MCP-aware client (Claude Desktop, Claude Code, Cursor, Zed, Windsurf, etc.) can connect to it, and the LLM gains live, searchable access to every page in this documentation — no copy-pasting required.

Server URL

https://docs.hyperclip.co/mcp

Tools the server exposes

search_hyperclip_docs

Semantic search over the Hyperclip docs knowledge base. The model uses this for broad or conceptual questions (“how do I authenticate”, “how is a run billed”, “how do I run an inline flow”). Returns contextual chunks with titles and direct links.

query_docs_filesystem_hyperclip_docs

Shell-style read-only filesystem rooted at the docs. The model can ls, tree, rg, cat, head, tail, grep, jq, etc. against every .mdx page and the OpenAPI specs. Useful for exact keyword matches, exploring structure, or pulling the full content of a specific page (e.g. cat /api/inputs.mdx).

Wiring it up

{
  "mcpServers": {
    "hyperclip-docs": {
      "url": "https://docs.hyperclip.co/mcp"
    }
  }
}

{
  "mcpServers": {
    "hyperclip-docs": {
      "url": "https://docs.hyperclip.co/mcp"
    }
  }
}

{
  "mcpServers": {
    "hyperclip-docs": {
      "url": "https://docs.hyperclip.co/mcp"
    }
  }
}

{
  "context_servers": {
    "hyperclip-docs": {
      "command": {
        "path": "npx",
        "args": ["-y", "mcp-remote", "https://docs.hyperclip.co/mcp"]
      }
    }
  }
}

After a restart, ask the model questions like:

“Look up how to handle the media_type_not_accepted error and write a recovery wrapper.”
“Read the inputs reference and generate a TypeScript type for the RunInputs shape.”
“Walk through the quickstart and produce a working Python script that polls until completion.”

The model will issue search_hyperclip_docs and query_docs_filesystem calls under the hood, then write code grounded in the actual documentation rather than guessing from training data.

The MCP server is read-only and stateless — it cannot run commands on your machine, write to your filesystem, or call the Hyperclip API. It only surfaces this documentation.

Option 3 — Manual context priming

If your tool doesn’t support MCP yet (or you want to seed a specific chat without configuring servers), paste the docs you need directly.

What to prime with

Different tasks need different context. Start narrow, expand only if the model needs more:

Task	Pages to paste
First integration	Quickstart, SDKs
Driving a flow with prompts/media	Inputs reference, Flows
Overriding voices, captions, models	Inputs reference — the `model_overrides` section
Inline `flow_schema` instead of saved flows	Flows § inline, Nodes
Picking a video / image / voice model	AI Models overview and the per-category pages
Error recovery	Errors
Pricing / cost estimation	Pricing

Quick recipe: prime Cursor / Claude Code / Copilot Chat

Open the docs page you need.
Click Raw or View source to grab the .mdx, or use Cmd-A then Cmd-C on the rendered page.
Paste into your chat with a framing line: “Reference the Hyperclip docs below. Then write a Node.js script that runs flow <id> with a prompt I’ll provide.”

That’s it — no special setup, just a focused context window.

For longer agentic tasks, paste the page once at the start of the session and tell the model “remember this is the source of truth for the Hyperclip API”. The model will refer back to it instead of inventing endpoint names.

Prompt patterns that work well

A few framings that produce consistently good Hyperclip integrations from a freshly-primed LLM:

Schema-first generation

“Given the FlowSchema for flow <id> (here’s the output of GET /flows/<id>), generate a strongly-typed input builder. For each required_inputs entry, accept the right value type and produce the correct inputs object.”The model uses the live schema as ground truth and produces a thin DSL over your specific flow — far better than asking it to write generic code.

Error-aware retry wrappers

“Read the error code list at /api/errors. Write a TypeScript helper that retries on transient errors (worker_error, render_failed, network timeouts) but not on terminal ones (invalid_flow_schema, monthly_cap_exceeded, insufficient_credits).”Asking the model to categorize errors against the catalog gives you a recovery policy that matches the actual API.

From-scratch SDK in a new language

“Read the SDKs reference, runs, and flows. Generate a Go client with the same method shape as the Node SDK (runs.create, runs.get, runs.wait, flows.list, flows.get).”The MCP server makes this practical — the model can cat each endpoint page in one round trip instead of guessing the wire format from training data.

Diff-style upgrade guidance

“My code uses inputs.media[2].type = 'image'. Check the flow definition (GET /flows/<id>) and tell me whether that’s actually allowed, or whether I should be sending 'video' based on accepted_media_types.”The eager validation API (media_type_not_accepted errors) is meant to catch this at runtime, but having an LLM check it at authoring time is even better.

Why this matters

The Hyperclip API is intentionally schema-driven — every flow has a discoverable shape (required_inputs, accepted_media_types, validation errors with structured codes). LLMs do well with schema-driven APIs because they can introspect, reason, and validate without inventing. Pair an MCP-aware coding tool with the eager validation in POST /runs and you get a tight loop: the model writes code based on the live schema, the API rejects bad calls with a specific error code, the model reads the error catalog and self-corrects. In practice this gets you from blank file to working integration in one prompt.

Where next

SDKs

Official Node and Python clients — the easiest target for an LLM-written integration.

Inputs reference

Every shape an LLM needs to understand to drive a run.

Errors

The structured catalog to point your retry/recovery prompts at.

Quickstart

The first thing to paste into a fresh chat session.

Introduction

Endpoints

Reference

Examples

AI tooling

Building with AI coding tools

Option 1 — One-click: copy the whole docs to your LLM

Suggested first prompt

Option 2 — Hosted MCP server

Server URL

Tools the server exposes

Wiring it up

Option 3 — Manual context priming

What to prime with

Quick recipe: prime Cursor / Claude Code / Copilot Chat

Prompt patterns that work well

Why this matters

Where next

SDKs

Inputs reference

Errors

Quickstart

​Option 1 — One-click: copy the whole docs to your LLM

​Suggested first prompt

​Option 2 — Hosted MCP server

​Server URL

​Tools the server exposes

​Wiring it up

​Option 3 — Manual context priming

​What to prime with

​Quick recipe: prime Cursor / Claude Code / Copilot Chat

​Prompt patterns that work well

​Why this matters

​Where next

SDKs

Inputs reference

Errors

Quickstart

Option 1 — One-click: copy the whole docs to your LLM

Suggested first prompt

Option 2 — Hosted MCP server

Server URL

Tools the server exposes

Wiring it up

Option 3 — Manual context priming

What to prime with

Quick recipe: prime Cursor / Claude Code / Copilot Chat

Prompt patterns that work well

Why this matters

Where next