First route the agent sees

One card opens into inspect, search, schema, and call only when needed.

osv / inspect / search_tools / get_tool / call_tool

npm install -g caplets

Requires Node 24+. Run setup from the install section next.

Why Caplets

Agents do not need every tool at once.

Direct MCP flattens provider APIs into the prompt. The agent spends context reading tool names, giant schemas, and setup assumptions before it can do the work.

Caplets keeps the first surface small, then opens the exact operation path when the agent asks.

discover cards

inspect tools

call typed

Benchmark proof

The result survived the surface cut.

Same task set, same 10/10 completion. Code Mode kept the agent focused by cutting the prompt surface instead of flattening every tool into view.

Run June 2026 with the real-world large MCP suite, openai-codex/gpt-5.5, 10 tasks, 2 runs per task, and a large no-fixture MCP stack.

The live run rows are listed here; the linked benchmark document covers the harness, deterministic surface check, and reproduction path.

tasks cleared

10/10

Caplets Code Mode, progressive modes, direct MCP, and Executor.sh all completed the task set.

avg tokens

236,803

Request plus output estimate for the Code Mode run, averaged across live Pi evals.

vs vanilla

72.0% fewer

Reduction against direct vanilla MCP without giving up the completed-task result.

Code Mode used 72.0% fewer request + output tokens than Direct vanilla MCP and 65.0% fewer than Executor.sh. Progressive disclosure also reduced tokens while every listed mode completed 10/10 tasks.