preview_snapshot has no scoping option and costs ~17k tokens plus a PNG per call

[nitrimandylis]

preview_snapshot returns page state, semantic elements, diagnostics, action history and a PNG screenshot on every call, and its only parameter is tabId. There is no way to ask for less.

On a plan-limited Claude subscription that is expensive, and the cost is not obvious from the tool description.

Measurements

I parsed my local Claude Code transcripts (~/.claude/projects/**/*.jsonl) and measured the text length of every tool result, grouped by tool. Text blocks only, so the PNG is on top of every number below.

Tool	Calls	Avg chars	~Avg tokens
mcp__t3-code__preview_snapshot	13	69,634	~17,400
playwright/browser_snapshot	16	7,326	~1,830
Read	2105	4,359	~1,089
Bash	8368	1,215	~303

About 9.5x the equivalent Playwright snapshot and 57x a shell result.

The larger effect is that a tool result stays in the conversation and is re-sent on every subsequent request in that thread. Weighting tokens the way billing does (cache read x1, cache write x2, output x5), deduplicating by requestId, and attributing each preview_* result to every later request in the same session:

Session	Requests	preview calls	Their tokens	Re-read cost	Share of session
A (UI work)	173	46	80,648	9.7M	26.7%
B (UI work)	58	21	61,178	1.5M	16.0%
C (light preview use)	155	2	23,499	2.0M	6.0%

Session A spent over a quarter of its entire token budget re-reading snapshots.

As an independent check, Claude Code's own /usage breakdown attributed 16% of my last 24 hours to the t3-code MCP server, against 4% for playwright over the same window.

What would fix it

Any one of these, roughly in order of usefulness to me:

1. A screenshot-only tool. preview_snapshot is currently the only way to get a PNG, so "let me look at the page" forces the full text dump along with it. Most of the time the image alone answers the question.
2. A verbosity or scope parameter on preview_snapshot, e.g. include: ["semantics"] / ["screenshot"] / ["diagnostics"], or a CSS/role selector to snapshot a subtree rather than the document.
3. Drop action history from the default payload. It grows monotonically within a tab, so repeated snapshots in one thread each carry a longer copy of the same log.

Happy to share the parsing script if the numbers are useful to reproduce.