PacketSmith — documentation¶
Everything here is written against code that exists. When a number appears — a latency, a token count, a contrast ratio — it was measured, and the page says how it was measured.
Using it¶
| Getting started | install, connect a provider, first build |
| Commands | everything behind /, and how the palette works |
| Providers and plans | the two-level catalog, the three wire protocols, usage meters, discovery |
| Themes and contrast | roles, the contrast test, writing your own |
| Troubleshooting | the failures that are not obvious |
How it works¶
| Architecture | the two engine classes, the MCP, and why the panel is derived rather than dictated |
| The system prompt | what the agent is told, what it is deliberately not told |
| Development | tests, conventions, where to add things |
For instructions aimed at AI agents working on this repository, see AGENTS.md — it carries the traps and the measurements, not the tutorials.
The three ideas¶
Nothing that changes is written down. The MCP server is asked what it can do — its 61 tools arrive with their JSON Schema attached, and tool 62 shows up on its own. Model lists come from models.dev. Providers are discovered from the same catalog. A list typed into source is stale the week after you type it.
A provider is not an endpoint. Kimi is one provider with two plans — a coding subscription and a metered API — with different URLs, protocols, prices and keys. Listing them as two providers is the kind of small lie that makes you paste the wrong key and get a 401 that explains nothing.
The panel shows what happened, not what was claimed. It is derived from tool results, so a device the model says it created but did not, does not appear.