RELAYS
agent handoff consoleBrief
What should the relay build?
The Architect turns this into a system plan. Everything downstream is derived from it.
Routing
What each agent needs from a model
Reasoning depth over speed. It runs once per relay and everything downstream inherits its mistakes, so this is the place to spend. Long context helps when you supply documents.
Structure and instruction-following. It has to hold the whole plan and emit consistent task ids and dependencies. A mid-tier model is usually enough.
Code quality and a large output budget — it emits whole files. Coding-tuned models earn their place here. Raise its max output tokens before anything else.
Judgement, not generation. It reads a plan and a build and returns a short verdict, so it can be the same model as the Architect or a cheaper one.
Route
How calls leave the browser
Hosted providers go through your Cloudflare function so keys can live in server env vars. Local endpoints are called direct — the browser must reach them.
Checking whether this deployment has functions…
Repeatability
WebGPU
Models run entirely in this tab. Nothing leaves the device.
MCP
Connect an external tool server
Streamable HTTP servers are reached through your Cloudflare function, which handles CORS and holds the auth header. On a static deployment there is no function, so the browser calls the server directly and it must send CORS headers. Tools discovered here are offered to whichever agents you allow.
Built in
Tools that need no server
Always available to the relay. These run in the page or through your own function.
AI Document
Context the relay can draw on
This is an input, not an agent. Nothing here generates output of its own — it supplies grounding to the other three. Text is chunked and scored with BM25 in the browser, and the highest-scoring passages are pasted into the Architect, Tech Lead and Developer prompts. No embeddings, no upload, works offline.
Retrieval is keyword-based, so the brief and the documents have to share vocabulary. Every stage reports which passages it received, or why it received none.
Binary formats such as PDF and DOCX are not parsed here. Paste their text instead.
Build output
No files yetHistory
Runs are stored in this browser only. Token counts come from provider responses; where a provider omits them the cell reads “not reported” rather than a guess.