Usage & tools

Once connected, Sankshep exposes these tools to your MCP client's model. In Copilot that means Agent mode; in Claude clients they're available directly.

Tools

Tool	What it does
get_context	Token-minimized context for given paths, relevance-ranked (semantic + lexical): strips comments, collapses method bodies that aren't relevant to your query, packs under a token budget, and reports the savings. Returns everything as one plain, readable text block — a short header stating how much was compressed and what was withheld (so nothing is silently truncated), followed by the code.
search_code	Semantic (nearest-neighbor) search over a local embedding index (bge-small, ONNX, offline).
index_repo	Builds/refreshes the semantic index over code and .docx / .pdf documents.
summarize_repo	Tree-sitter-backed API surface of a tree, in every language Sankshep parses. Bounded by maxTokens, and it says so when it truncates.
remember / recall	A per-repo, branch-scoped fact store — decisions and conventions that persist across sessions and clients.
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.

Prompts

Prompt	What it does
compose_task_prompt	Assembles a grounded, one-shot prompt for a coding task from minimized code (get_context) + recalled conventions (recall). A prompt, not an answer. See the composer guide.

MCP clients that surface prompts (Claude Code, Claude Desktop) expose compose_task_prompt as a slash-command.

Supported languages

Sankshep parses code with tree-sitter. Every language below is pinned by a golden test — Sankshep claims only what it verifies.

Language	Comment strip	Body collapse	Notes
C#	✓	✓	Trims unused usings at level=Aggressive.
JavaScript / TypeScript	✓	✓	.js .mjs .cjs .jsx .ts; .tsx uses its own grammar, so JSX inside a component is parsed rather than guessed at
Go	✓	✓
Java	✓	✓
C / C++	✓	✓	.c → C; .cpp .cc .cxx .h .hpp .hh .hxx → C++
Rust	✓	✓
PHP	✓	✓
Python	✓	—	Indentation blocks have no brace body to elide.
Ruby	✓	—	def…end blocks have no brace body to elide.

Body collapse replaces a method/function body that isn't relevant to your query with a { /* … body elided (N lines) */ } marker, so it applies only to brace-bodied languages; Python and Ruby get comment-stripping and packing but keep their bodies. Files in any other language are still retrieved, ranked, and packed under your token budget — just at the text level, without parse-aware minimization.

Minimization levels

get_context takes a level argument (see the tool reference). Conservative collapses nothing; Balanced (the default) keeps the method bodies relevant to your query and collapses the rest; Aggressive collapses every non-relevant body plus expression-bodied (=>) members and trims unused usings.

A typical flow

1. Index once: ask the assistant to run index_repo on your repo (or a subtree). This embeds code + documents locally.
2. Ask grounded questions: the model calls get_context / search_code to pull only the relevant, minimized slice — not whole files.
3. Remember what you decided: "Remember: we sign JWTs with RS256, keys rotate monthly." The fact is stored locally and recalled later, even from a different client.
4. Compose a task: run compose_task_prompt to get a grounded, structured prompt for a change.

Category matters for step 4

compose_task_prompt's "Project conventions" section injects facts remembered under the convention category and no other. File a rule you want the model to follow as convention; decision is for the record and for export_decisions. The composed prompt names the categories it skipped, so an empty section tells you where your facts went.

Cross-client memory

Because facts live in a local database every client reads, a decision recorded in Claude Code can be recalled next week in Copilot Chat — cross-client shared memory is an emergent benefit of keeping everything local. See How it works.

Measuring the savings

token_report returns cumulative token accounting per tool: compression — delivered tokens against the original size of the same files, i.e. what minimization actually removed — plus how much context was searched to find them.

Searched tokens are reported for context and are deliberately not a baseline: Sankshep may read a whole repository to select a handful of files, and that is not context you would otherwise have sent. Counting it as a saving is how a tool ends up claiming it saved more tokens than the model could ever accept. The value of selecting the right files is a retrieval question, and recall — not a token count — is what answers it.

There is no dollar figure. Sankshep is never told which model you use or what you pay for it, so any total would be an assumption dressed as a measurement. Multiply compressed tokens by your own negotiated input rate if you want one. Methodology: Benchmarks.