AI-first workflow manager for heavy MCP tools

Heavy MCP tools don’t belong in your prompt.

ToolCapsule inventories existing MCPs, turns selected heavy servers into lazy-loaded, file-first Agent Skills, and keeps calls patchable and retryable.

Install the ToolCapsule Skill npx skills add RainSunMe/toolcapsule --skill toolcapsule Then ask your agent to run tcap init --url .

Agent onboarding prompt

Copy a detailed prompt for Claude Code, GitHub Copilot / VS Code, OpenCode, Gemini CLI, or Cursor so it can import your MCP setup safely.

View on GitHub

88.4% fewer input tokens

87.6% fewer total tokens

100% planning success rate

The problem

MCP made tools universal. It also made tool context heavier.

Native MCP is excellent for small, frequent tools. But document and workflow MCP servers often expose long descriptions, schemas, and rules that can be visible to the model even when the user is only planning.

ToolCapsule keeps MCP as the capability layer and uses Skills as the agent-facing workflow layer. Full schemas are still available, but no longer need to be carried every turn.

MCP-to-Skill conversion

Turn existing MCP config into a lazy-loaded Agent Skill.

ToolCapsule reads MCP registrations you already have and writes a profile plus an Agent Skill for Claude Code, GitHub Copilot / VS Code, OpenCode, Gemini CLI, or Cursor.

Claude Code VS Code OpenCode Gemini CLI Cursor

toolcapsule mcp

$ tcap init  --url 
STATUS  NAME    SCOPE      SOURCE  MODE    PATH
● on    github  workspace  vscode  native  .vscode/mcp.json
● on    notion  workspace  claude  native  .mcp.json

$ tcap init github --target claude
created  ~/.toolcapsule/profiles/github.json
created  .claude/skills/github-mcp/SKILL.md

$ tcap tools github --brief

No retyping

Reuse existing URLs, commands, headers, env references, and cwd.

Safe by default

Workspace config first; user-level config requires opt-in.

Agent-ready

Default target is Claude, with Copilot, OpenCode, .agents, or all targets available.

The capsule

A tool call becomes a patchable artifact.

Instead of asking the model to regenerate a large tool call after a failure, ToolCapsule stores the arguments, payloads, request, response, and error as local files.

runs/2026-06-06T10-00-00

args.json
content.md
request.json
response.json
error.txt
command.txt

Discover briefly

List MCP tools as compact summaries instead of injecting every schema.

Write locally

Put Markdown, JSON, and long arguments into files the agent can patch.

Call deterministically

Run the MCP tool through a CLI and save the full request/response trace.

Patch and retry

Fix the local file and replay the same tool call without rewriting the prompt.

Native MCP

Regenerate the call

Tool schemas stay close to the prompt.
Large payloads are reassembled by the model.
Retries can accidentally change correct content.

ToolCapsule

Patch the artifact

Schemas are lazy-loaded when needed.
Payloads live in args.json and content files.
Retries are auditable, local, and deterministic.

For heavy MCP servers

Keep the protocol. Change the workflow.

ToolCapsule is designed for Feishu/Lark, Notion, Jira, Linear, Confluence, GitHub management tools, and internal SaaS MCPs.

Start with ToolCapsule