Agent DX

FIG.

FIG. 00 · AGENT DXMCP + SKILLS

Your coding agent already knows Synapse Garden. Wire it in 60 seconds. Open Claude Code, Cursor, Windsurf, Cline, or Codex; install our MCP server and our skill pack; ask the agent to "wire Synapse Garden into this project". A minute later you have a scoped key, a working stream chat call, and a budget cap — without typing a URL or scrolling these docs.

FIG. 01TWO SURFACES

SCHEMATIC

The two surfaces

How they compose

The MCP server gives the agent live state about your account: which models are available, which keys exist, what the last 100 errors look like, where this month's spend went. The skill pack gives the agent how-to knowledge: idiomatic AI SDK code, migration diffs from the OpenAI and Anthropic SDKs, debug recipes for common 4xx responses. Together they replace the moment a developer pastes a URL into a chat: the agent reads mg://catalog, picks a model, calls create_api_key, writes the env var, and edits the file — all without leaving the editor.

Concrete flows

A few examples of what an agent can do once both surfaces are wired:

• Wire Synapse Garden into this Next.js app. The agent reads mg://catalog, picks openai/gpt-4o-mini, calls create_api_key, writes MG_KEY into .env.local, edits app/page.tsx to call the AI SDK against https://synapse.garden/api/v1, runs npm run dev, and confirms a 200 response. No tab switching.
• My bill looks suspicious. The agent calls analyze_spend, summarizes the top model and project pairs by day, flags an anomaly, and offers to call set_project_budget on the offending project. You approve the confirmation, the budget lands, and the audit trail records who did it.
• Migrate this file from the OpenAI SDK direct. The agent calls migration_plan on the open file, applies the two-line diff (baseURL + apiKey), prefixes model ids with openai/, runs the file, and tails errors with tail_errors to confirm a clean cutover.

What's next