Skill pack
A curated bundle of Anthropic Agent Skills that teach your coding agent how to use Synapse Garden idiomatically. AI SDK recipes, migration playbooks, debugging.
@synapse-garden/skills is a pack of Anthropic Agent Skills — small markdown bundles that teach your coding agent how to use Synapse Garden the way an experienced engineer would. Each skill is a directory with a SKILL.md file (loaded into context by default) and a set of companion recipes/*.ts files (loaded on demand). The agent picks the right skill from the description hook, reads only the recipe it needs, and stays cheap on tokens even when the body is large.
Install
One command for Claude Code:
claude-code skill add @synapse-garden/skillsOr copy the skills into your project manually:
npx @synapse-garden/skills extract --to .claude/skills/The same source is reachable as a hosted skill manifest at https://skills.synapse.garden/<name> for clients that consume the Vercel Skill Creator pattern. Cursor and Windsurf get matching .cursor/rules/ and .windsurf/rules/ adapters generated from the same skill bodies in CI — single source of truth, multiple distribution targets.
Skills shipping in v1
Eight skills. The first three ship in wave 1 — they cover the path most projects take: install, write idiomatic AI SDK code, debug when something goes wrong. The remaining five follow.
| Skill | Wave | When the agent loads it | Body content |
|---|---|---|---|
synapse-garden-quickstart | 1 | "set up Synapse Garden", "wire in MG", "first call" | Three install variants (AI SDK, OpenAI SDK, Anthropic SDK), env vars, smoke-test snippet, link to MCP install. |
synapse-garden-ai-sdk | 1 | The agent is editing AI SDK code (streamText, generateObject, embed, tools) | Code recipes for streamText, generateObject, embed, tool use, vision, and createOpenAICompatible factory wiring. |
synapse-garden-debug | 1 | "Synapse Garden is returning 4xx", "why is X failing" | Error code map, x-mg-trace-id correlation, MCP tail_errors link, log-and-retry recipe. |
synapse-garden-migrate-openai | next | The agent is editing a file that imports openai | Side-by-side before and after, env var swap, gotchas (model-name prefix, OpenAI-Organization ignored). |
synapse-garden-migrate-anthropic | next | The agent is editing a file that imports @anthropic-ai/sdk | Side-by-side, base URL swap, content blocks unchanged. |
synapse-garden-rag | next | "build a RAG pipeline", "search my docs" | Full embed → ANN → rerank → generate recipe with a pgvector schema and embedMany batch loop. |
synapse-garden-tools | next | "add tool use", "function calling" | Zod tool definitions, maxSteps, multi-step debug pattern. |
synapse-garden-vision | next | "send an image to the model" | Base64 vs URL, detail levels, multi-image patterns. |
Each skill stays under 300 lines of SKILL.md plus its recipes. If a skill grows past that, we split it.
Authoring your own
The skill format is open — once you've installed @synapse-garden/skills, the same authoring conventions apply to skills you ship for your own team. See Anthropic's Agent Skills reference for the spec, and Vercel's Skill Creator for the recipe-led authoring pattern we mirror. A guide for shipping team-private skills that compose with @synapse-garden/skills and @synapse-garden/mcp lives at /docs/agents/authoring.
What SKILL.md looks like
Every skill follows the same shape: YAML frontmatter with name and description (the description is the agent's "should I use this?" hook — keep it under 200 chars and write it for the agent's mental model, not the user's), a short body of section headings, and pointers into recipes/*.ts files. Here's the head of synapse-garden-ai-sdk:
---
name: synapse-garden-ai-sdk
description: Use when the user is writing or editing AI SDK code (streamText, generateObject, generateText, embed, tools) and the project uses Synapse Garden. Provides idiomatic recipes that wire the AI SDK to Synapse Garden's OpenAI-compatible endpoint.
---
# AI SDK with Synapse Garden
Synapse Garden speaks OpenAI-compatible on `/v1/chat/completions` and Anthropic-
compatible on `/v1/messages`. Use the AI SDK's `createOpenAICompatible`
factory once, reuse the resulting client everywhere.
## One-time setup
See `recipes/setup.ts`.
## Streaming text
See `recipes/stream-text.ts`.The body is essentially a table of contents pointing into recipes/*.ts. The agent reads only the file it needs — the rest stays on disk until something else asks for it.
Compose with the MCP
The same recipes are also exposed as MCP prompts in @synapse-garden/mcp. Vanilla MCP clients without skill support — including agents that consume the protocol's prompt surface but not Anthropic's skill format — still get the recipe surface, just routed through a different transport. One source of truth, two ways to reach it.
The skills/ tree is canonical. The npm package is a tarball of that tree. The Cursor and Windsurf adapters are generated in CI. The MCP prompts surface reads the same source. When a recipe changes, every target picks it up on the next release — no parallel maintenance.
MCP server
Connect any MCP-capable agent (Claude Code, Cursor, Windsurf, Cline, Codex) to Synapse Garden's control plane. Read live catalog, query logs, mint scoped keys.
Setup flow · Claude Code
From a fresh repo to a streaming chat call in under a minute, driven entirely by your coding agent. No tab switching, no copying URLs.