Skip to content

skel-mcp v0.2.0

Choose a tag to compare

@encryptedbasic encryptedbasic released this 14 Nov 13:29
· 0 commits to 858412ed15fbc34d2f9c7c2fde0c610d1b5ddc05 since this release

skel-mcp v0.2.0

Get the code an agent needs in one MCP call. This release makes get_context the default interface for TypeScript, JavaScript, Python, and Go.

Launch directly from this GitHub Release archive.

npx --yes --package=https://github.com/encryptedbasic/skel-mcp/releases/download/v0.2.0/skel-mcp-0.2.0.tgz skel-mcp

Install from the npm registry with:

npx --yes skel-mcp@0.2.0

Use Node.js 20 or later. Install Git to use the diff operation. The server runs locally over stdio. See the setup guide for Claude Code, Cursor, Codex, and VS Code.

What's new

  • get_context batches up to 20 requests for names, signatures, full symbols, selected lines, or AST-aware diffs.
  • Symbol responses include relevant imports and related type declarations. TypeScript and JavaScript can follow an unambiguous relative import one step within the workspace. Python and Go resolve declarations in the source file.
  • Responses start with a 1,200-token page and include a continuation cursor when more content is available. Set maxTokens: null for the full response. Use detail: "full" on a diff request for complete affected implementations.
  • Source reads are scoped to the configured workspace. The server rejects paths outside the project and paths to Git metadata.
  • Start the server with --legacy to use get_file_skeleton, get_symbol, and get_smart_diff.

Benchmark

Codex CLI benchmark token totals

Scenario Without MCP With skel-mcp Fewer tokens
Six separate tasks 212,180 149,252 29.7%
The same six tasks in one batch 41,966 26,216 37.5%

The saved experiment measured total input plus output tokens in Codex CLI with direct MCP tool routing. It ran one paired trial per scenario, and all reference answers matched. The baseline used full-file reads or ordinary Git diffs. Both arms of the batch experiment retrieved their context in one call. Results depend on the workload and the client's tool configuration. Six-task report · Batch report