Provides five information retrieval scoring tools for evaluating RAG system quality: recall@k and hit@k measure whether relevant documents appear in your top results, MRR scores the rank of the first correct match, NDCG@k weights graded relevance scores by position, and evaluate_batch runs all metrics on multiple queries at once. Complements the author's promptbudget, citecite, and ragdrift servers from the same mcp-stack collection. Install it when you need Claude to help you benchmark retriever performance against ground truth relevance judgments, either interactively or as part of a larger RAG development workflow. The logic is implemented in plain JavaScript so you can run it via npx without pulling in the Rust ragmetric crate.
Fourteen small MCP servers for everyday LLM / RAG / agent work. Each ships independently on npm, listed in the official MCP Registry, and is callable from Claude Desktop, Cursor, Cline, Windsurf, Zed, and any other MCP-compatible client.
| Package | Tools | When to install |
|---|---|---|
@mukundakatta/promptbudget-mcp | truncate_to_token_budget | LLM needs to fit text into a context window mid-task |
@mukundakatta/citecite-mcp | inject_citations, parse_citations, strip_citations | Round-trip [1] [2] markers in RAG outputs |
@mukundakatta/ragmetric-mcp | recall_at_k, hit_at_k, mrr, ndcg_at_k, evaluate_batch | LLM helps score retrieval quality |
@mukundakatta/ragdrift-mcp | interpret_drift_score, recommend_thresholds, explain_drift_dimensions | LLM diagnoses RAG drift alerts |
| Package | Tools | When to install |
|---|---|---|
@mukundakatta/csv-tools-mcp | parse_csv, to_csv, pluck_columns | LLM needs RFC 4180-correct CSV (quoted fields, embedded commas, BOMs) |
@mukundakatta/regex-test-mcp | test_regex, find_all, replace | LLM needs exact regex semantics with real match offsets and named groups |
@mukundakatta/jmespath-mcp | json_query | LLM needs deep JSON traversal (filters, projections, pipes) without hallucinating |
@mukundakatta/diff-mcp | unified_diff, apply_patch, parse_patch | Code-review or code-edit agent needs character-precise patches |
@mukundakatta/sqlfmt-mcp | format_sql, list_dialects | Deterministic SQL formatting across 19 dialects |
@mukundakatta/shellquote-mcp | quote_bash, quote_bash_argv, quote_cmd, quote_powershell | Safe shell argument escaping; LLMs constantly mishandle quotes/$VAR |
@mukundakatta/json5-mcp | parse_json5, to_json5, to_strict_json | LLM emitted JSON-with-comments / trailing-commas; round-trip to strict JSON |
@mukundakatta/toml-yaml-json-mcp | parse, format, convert | TOML / YAML / JSON conversion (LLMs especially mishandle TOML) |
@mukundakatta/timezone-mcp | convert_tz, now_in, tz_offset | IANA timezone math with real DST rules; LLMs hallucinate offsets |
@mukundakatta/html-to-markdown-mcp | html_to_md, extract_text | Web-scraping agents need clean Markdown / text from HTML |
The first four wrap or re-implement logic from the user's Rust + Python work:
ragdrift (crates.io) /
ragdrift-py (PyPI) — five-dimensional drift detectionembedrank, promptbudget,
stopstream, citecite,
ragmetric — pure-Rust crates from
the rust-llm-stack workspaceThe MCP servers are independent: they implement the logic they need in plain JavaScript so installing them doesn't pull a Rust toolchain.
Add to your MCP client config. Example for Claude Desktop:
{
"mcpServers": {
"promptbudget": { "command": "npx", "args": ["-y", "@mukundakatta/promptbudget-mcp"] },
"citecite": { "command": "npx", "args": ["-y", "@mukundakatta/citecite-mcp"] },
"ragmetric": { "command": "npx", "args": ["-y", "@mukundakatta/ragmetric-mcp"] },
"ragdrift": { "command": "npx", "args": ["-y", "@mukundakatta/ragdrift-mcp"] },
"csv-tools": { "command": "npx", "args": ["-y", "@mukundakatta/csv-tools-mcp"] },
"regex-test": { "command": "npx", "args": ["-y", "@mukundakatta/regex-test-mcp"] },
"jmespath": { "command": "npx", "args": ["-y", "@mukundakatta/jmespath-mcp"] },
"diff": { "command": "npx", "args": ["-y", "@mukundakatta/diff-mcp"] },
"sqlfmt": { "command": "npx", "args": ["-y", "@mukundakatta/sqlfmt-mcp"] },
"shellquote": { "command": "npx", "args": ["-y", "@mukundakatta/shellquote-mcp"] },
"json5": { "command": "npx", "args": ["-y", "@mukundakatta/json5-mcp"] },
"toml-yaml-json": { "command": "npx", "args": ["-y", "@mukundakatta/toml-yaml-json-mcp"] },
"timezone": { "command": "npx", "args": ["-y", "@mukundakatta/timezone-mcp"] },
"html-to-md": { "command": "npx", "args": ["-y", "@mukundakatta/html-to-markdown-mcp"] }
}
}
git clone https://github.com/MukundaKatta/mcp-stack
cd mcp-stack
npm install
npm test --workspaces
MIT OR Apache-2.0.
io.github.ericm1018/skillfm-llm-cost-optimizer-openai-anthropic-usage
io.github.mikerawsonnz/llm-orchestration-agent
io.github.mikerawsonnz/authenticated-llm-agent
labforgedev/copilot-memory-mcp
csoai-org/agent-prompt-injection-firewall-mcp
io.github.mikerawsonnz/authenticated-multi-llm-agent