sankshep.linux-x64 3.0.0

<p align="center"> <img src="https://raw.githubusercontent.com/nitinpawar28/sankshep-docs/main/docs/assets/logo-stacked.png" alt="Sankshep" width="280"> </p>

Sankshep (संक्षेप)

Sankshep is Sanskrit for "the concise essence" — a summary that keeps the substance and drops the excess.

Maximum context, minimum tokens — with the benchmarks to prove it (see the numbers).

Sankshep is a proprietary MCP server, written in C#/.NET 10, free to install and use, that gives MCP clients — VS Code Copilot Chat, Claude Code, Claude Desktop, Cursor — token-minimized, relevance-ranked codebase context, plus a persistent, per-repo memory of facts and decisions. Everything runs locally: no code or content leaves your machine, and there's no telemetry by default.

Install

dotnet tool install -g sankshep
sankshep serve --repo .

The tool bundles the tree-sitter, sqlite-vec, and ONNX native assets for your platform. Sankshep is also available as a container image (ghcr.io/nitinpawar28/sankshep) for HTTP / server deployments — see the deployment guide.

The server speaks MCP over stdio — point an MCP client at it rather than running it standalone in a terminal.

Connect your MCP client

Minimal VS Code config (.vscode/mcp.json), using the installed tool:

{
  "servers": {
    "sankshep": {
      "type": "stdio",
      "command": "sankshep",
      "args": ["serve", "--repo", "${workspaceFolder}"]
    }
  }
}

Per-client setup for Claude Code, Claude Desktop, and Cursor is in the install guide.

What it does

Sankshep shrinks the code and docs an AI assistant reads, without dropping the parts that matter:

• get_context — token-minimized, ranked context for given paths: strips comments, collapses non-target method bodies, packs under a token budget, and reports the savings.
• search_code — pure semantic (nearest-neighbour) search over a local embedding index (bge-small, ONNX, runs offline). It does no lexical matching: for blended semantic + lexical ranking, pass a query to get_context.
• index_repo — builds the semantic index over code and .docx/.pdf documents.
• summarize_repo — tree-sitter-backed API surface of the repo, across every language Sankshep parses. C# reports namespaces and the public surface; the other languages list declared types and functions. A tree with no supported sources says so rather than returning an empty summary.
• remember / recall — a per-repo, branch-scoped fact store.
• export_decisions — writes remembered decisions to a DECISIONS.md.
• token_report — cumulative token accounting per tool: how far minimization compressed the code it delivered, and how much context it searched to find it.
• compose_task_prompt (an MCP prompt, surfaced as a slash-command) — assembles a grounded, one-shot prompt for a coding task from minimized code + recalled conventions. A prompt, not an answer. (composer guide)

Everything runs locally: embeddings via ONNX Runtime, vector search via sqlite-vec (with a pure-C# fallback), no code or content leaves the machine.

Documentation

Full guides — install & usage, the prompt composer, deployment (Docker / Windows Service / systemd / Kubernetes / air-gap), benchmarks, and a curated architecture overview — live at nitinpawar28.github.io/sankshep-docs.

Feedback

Sankshep is a proprietary, solo-maintained product and does not accept external pull requests. Bug reports and feature suggestions are genuinely welcome via the public feedback channel: sankshep-docs issues.

Security issues are different — do not report them publicly. Use the private disclosure channel in SECURITY.md (report a vulnerability).

License

Proprietary — Sankshep is free to install and use (including commercially) but is not an open project: its source is not distributed, and redistribution / modification are not permitted. See the LICENSE for terms.