Give your agent Capabilities, not a tool wall
Turn sprawling MCP servers into focused capability cards. Your agent can start with one typed route, zoom into tools only when needed, and keep huge schemas out of the prompt until they matter.
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.
Direct MCP
Too many tools
Every downstream operation lands in the agent's first view.
With Caplets
The agent starts with a named capability and opens only the route it needs.
Direct MCP
Too much schema
Large tool definitions compete with the user's actual task for context.
With Caplets
Schemas stay behind inspect, search, and get_tool until they matter.
Direct MCP
Too much setup
Every agent repeats provider wiring, OAuth, secrets, and MCP config.
With Caplets
One Caplets surface can be reused locally or from a remote server.
Setup
Start with the smallest useful Caplet.
caplets setup wires the agent integrations you choose. Add OSV first because it needs no auth; bring in GitHub or Sourcegraph after the discovery path feels right.
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.
Read method & reproduce- tasks cleared
- 10/10
- avg tokens
- 236,803
- vs vanilla
- 72.0% fewer
Caplets Code Mode, progressive modes, direct MCP, and Executor.sh all completed the task set.
Request plus output estimate for the Code Mode run, averaged across live Pi evals.
Reduction against direct vanilla MCP without giving up the completed-task result.
flat tool wall 215
first screen cards 7
surface cut 79.9%
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.
Remote Caplets server
One auth home. Every agent gets the same tools.
Run Caplets as a small HTTP service. Provider tokens and OAuth state stay server-side; Codex, OpenCode, Pi, Claude Code, and any MCP client attach to the same capability surface.
- MCP
- /caplets/v1/mcp
- Attach
- /caplets/v1/attach
- Admin
- /caplets/v1/admin
- Health
- /caplets/v1/healthz
Serve once
caplets daemon install --start
caplets remote host approve <code> --yesUse remotely
caplets remote login <url>
caplets attach <url>