Tokenminning in Zed
Zed is a GPU-rendered IDE with a built-in AI assistant. Each turn resends your prompt, project rules, @ attachments, and thread history. Most spend comes from long threads, frontier defaults, and broad context—not one verbose reply.
Hub for the Zed cluster. For the general stack see Where to start.
Quick checklist
- Check provider billing — token usage.
- Use fast/mid-tier models for routine work; frontier last.
- Trim rules and
@context — keep attachments narrow. - Wire local models (Ollama, llama.cpp) for exploration when quality allows.
- New assistant thread per task.
Typical impact: 40–65% routing routine work to mid-tier; 20–40% trimming rules and attachments; 30–50% less growth from shorter threads.
How Zed bills
Zed assistant uses your configured provider (Anthropic, OpenAI, hosted Zed inference, or local). No separate agent vs tab pools like Cursor — see Token usage.
Each turn sends: prompt + @ context, project rules, prior messages, tool results when agent workflows run.
Guides
| Guide | What you’ll learn |
|---|---|
| Token usage | Provider dashboards, input vs output diagnosis |
| Local models | Ollama, llama.cpp, vLLM, LiteLLM in Zed |
| Context & rules | Rules files, @ mentions, RAG-style retrieval |
Model routing
Default cheap, escalate on failure — Model routing:
| Tier | Use for |
|---|---|
| Inline / fast | Completions, single-line edits |
| Mid-tier assistant | Most multi-file questions |
| Frontier | Deep debug, architecture — last resort |
Related
Last updated on