MCP 服务器

contextsliver

An MCP server that indexes codebases into a local graph and provides on-demand context retrieval for AI coding agents, reducing token usage by tracking session history and delivering only relevant code subgraphs.

README

ContextSliver

On-demand context management for AI coding agents. Stop reading whole files — ask for the connected subgraph instead.

The problem

When you ask an AI coding agent (Claude Code, Cursor, Cline) to "fix the bug in AuthService," it repeatedly reads entire files to find the 5% that's relevant — burning 40,000–80,000 tokens on a question that needed ~3,000.

find . -name "*.ts"   → 2,000 tokens (a file listing)
cat AuthService.ts    → 3,000 tokens (whole file)
grep -r "AuthService" → 5,000 tokens (40 matches)
cat AuthMiddleware.ts → 2,500 tokens (whole file, again)
... × 10 more ...

Existing tools don't fully fix it: Graphify dumps one enormous whole-repo map into context up front; Repomix packs the entire repo into one file; Aider's repo-map sends a whole-repo summary with every message. None of them track what the agent has already seen this session.

What ContextSliver does

ContextSliver runs as a background MCP server on your machine. It indexes your codebase into a local SQLite graph of every function, class, and import. When the agent needs context, it calls an MCP tool instead of reading files:

Agent:   "What connects to AuthService? Budget: 2,000 tokens."

ContextSliver:
  symbol:        AuthService  (src/auth/AuthService.ts)
  callers:       [AuthMiddleware, LoginController]      ← who uses it
  dependencies:  [UserRepository, TokenService]         ← what it uses
  already_in_context: [UserRepository]                  ← skipped, agent already has it
  // ~380 tokens

Three things make it different:

On-demand pruning — never sends the whole graph, only the connected subgraph for the task.
Session ledger — tracks what the agent has already seen and skips re-sending it.
One-command setup — npx contextsliver init. No database server, no API key.

Quickstart

# In your project root:
npx contextsliver init      # creates .sliver/, .mcp.json, CLAUDE.md, indexes the repo
npx contextsliver start     # runs the MCP server + file watcher (stdio)

Then restart Claude Code / Cursor / Cline — they'll pick up the five tools automatically. See the templates for client-specific config.

The five MCP tools

Tool	What it does	Typical tokens
`cs_index_repo`	Trigger a full re-index	~50
`cs_get_context`	Symbol definition + immediate connections	~300–800
`cs_blast_radius`	All callers + dependents up to N hops	~500–2,000
`cs_search_symbols`	Full-text search across indexed symbols	~200–600
`cs_index_status`	Index health, file count, last-updated	~100

Pass the session_id from your first cs_get_context call to every subsequent call to enable deduplication.

Supported languages

TypeScript / JavaScript / TSX (v0.1)
Python (v0.1)
Go, Rust, Java — planned (see roadmap)

Adding a language = add a grammar package + a grammars/<lang>/tags.scm query + a fixture. See CONTRIBUTING.md.

How it works

Your codebase  ──chokidar──▶  Parser (Tree-sitter)  ──▶  SQLite graph (.sliver/index.db)
                                                                    │
                              MCP server (stdio) ◀──────────────────┘
                                    │   session ledger (.sliver/sessions.db)
                                    ▼
                          Claude Code / Cursor / Cline

Parser: Tree-sitter extracts symbols + imports per file.
Graph engine: stores symbol→symbol edges; bidirectional BFS (blastRadius) for blast radius with cycle detection.
Session manager: per-session ledger computes deltas so already-sent context is skipped.
MCP server: exposes the five tools over stdio.

Token counting

Counts use gpt-tokenizer (cl100k_base) and are labeled ~approximate — close enough for budget guidance, not billing.

Development

npm install
npm test            # unit + integration tests
npm run test:bench  # indexing benchmarks
npm run build       # tsc → dist/
npm run lint        # eslint

Requires Node ≥ 20.

Roadmap

v0.1 ✅ TS/JS + Python, SQLite graph, session ledger, 5 tools, CLI, watcher
v0.2 — incremental indexing polish, Cursor integration, CI benchmarks
v0.3 — Go + Rust, monorepo workspace resolution, language-plugin docs
v0.4 — PreToolUse hook, Java, published token-reduction benchmarks
v0.5 — Streamable HTTP transport, DuckDB backend for 50k-file repos, PageRank ranking
v1.0 — frozen API, optional native (napi-rs) engine, SCIP/LSP precision backend

See contextsliver-spec.md for the full specification.