Display Token Usage for Each Tool Call in Agent Panel

[l4b4r4b4b4]

Summary

Display token usage (input + output = total) for each tool call in Zed's agent panel. This would help users understand the cost and context consumption of individual tool operations, enabling better optimization and cost management when using AI features.

Motivation

The Problem:
Users have no visibility into how many tokens each tool call consumes. This makes it difficult to:

• Understand which tools are expensive (in tokens/cost)
• Optimize workflows for cost efficiency
• Debug why context windows fill up quickly
• Make informed decisions about tool usage

Why This Matters:

1. Cost Management — LLM API costs add up; knowing per-call usage helps budgeting
2. Context Optimization — Identify which tools return excessive data that could be trimmed
3. Debugging — Understand why context window fills up or why requests fail due to length
4. Tool Selection — Choose between similar tools based on token efficiency
5. Transparency — Extends Zed's philosophy of showing tool arguments (now implemented) to also show resource usage

Current vs. Desired Behavior

Current Behavior:

[Tool Call: read_file]
  path: "src/main.rs"
[Approve] [Decline]

Result: [file contents...]

→ No indication of token cost
→ User doesn't know this call consumed 2,000 tokens

Desired Behavior:

[Tool Call: read_file]
  path: "src/main.rs"
[Approve] [Decline]

Result: [file contents...]
📊 Tokens: 45 in → 1,955 out (2,000 total)

→ User immediately sees the cost
→ Can make informed decisions about future tool usage

Proposed UI Patterns

Option 1: Inline Token Count (Preferred)
Show a compact token summary after the result:

Result: {"matches": [...]}
📊 Tokens: 120 in → 340 out (460 total)

Option 2: Collapsible Details
Subtle indicator that expands on click:

Result: [10 matches found...]
📊 460 tokens ▼
   └─ Input: 120 | Output: 340

Option 3: Badge on Tool Call Header
Show tokens inline with the tool name:

[Tool Call: semantic_search] [460 tokens]
  query: "authentication middleware"

Option 4: Session Summary
In addition to per-call display, show cumulative usage:

Session: 12,450 / 128,000 tokens used (9.7%)

Use Cases

1. Cost-Conscious Development — Track API costs during development sessions
2. Large File Operations — See how much context read_file on large files consumes
3. Search Operations — Understand token cost of grep or semantic search results
4. Context Window Management — Know when approaching limits before hitting errors
5. Tool Comparison — Compare token efficiency between different approaches
6. Billing Reconciliation — Match Zed usage with provider billing

Technical Considerations

• Provider Support — Token counts come from API responses; availability varies by provider
• Tool-Level Granularity — Some APIs only report total usage, not per-tool; may need estimation
• Streaming Responses — Token count may only be available after completion
• Native vs MCP Tools — Should apply to both native Zed tools and MCP tools
• Performance — Displaying counts shouldn't add latency; use data already returned by APIs

Implementation Notes

Most LLM APIs (OpenAI, Anthropic, etc.) return usage objects with token counts:

{
  "usage": {
    "input_tokens": 120,
    "output_tokens": 340
  }
}

This data is already available in responses; the feature is primarily a UI presentation concern.

Related Features

This pairs well with existing and requested Zed features:

• ✅ Tool call arguments — Now visible before approval (Show tool call input arguments before approval in Agent Panel #46241)
• ✅ Context window display — Shows available context (Display both available and true context window for AI models #47041)
• ❓ MCP server name — Shows which server (related request)
• ❓ Per-tool-call tokens — This request

Together, these provide complete transparency for agent operations.

Additional Context

Discussion #45738 raised concerns about high token usage (7,000 tokens for a simple "hello"). While that discussion focused on reducing usage via prompt customization, this request focuses on visibility — showing users what's being consumed so they can make informed decisions.

Transparency about resource usage aligns with Zed's design philosophy of giving users control and understanding of their tools.

[itguy327]

This would be an excellent addition, along with custom system prompts

[UrRhb]

Great feature request. The billing reconciliation point is underrated — we've seen teams where the OpenAI dashboard shows $X/day but nobody knows which tool calls or features are driving that.

For anyone wanting this kind of per-request cost visibility outside of Zed (e.g. in your backend or API server), we built burn0 — it patches fetch at import time and shows per-call costs in the terminal. Zero config, works with OpenAI/Anthropic/50+ services. Not a replacement for what's being proposed here (in-editor UI would be amazing), but useful as a complement during development.

Re: the implementation — Option 2 (collapsible details) feels cleanest. The streaming concern is real though. Most providers only include usage in the final chunk, so the count can't be displayed until the response completes. For tool calls specifically, the usage field is reliable since the response isn't streamed.