AsterMind MCP
io.github.AsterMindAIPublished by GitHub account AsterMindAIv0.1.0Updated Oct 24, 2026
On-device reranking that cuts RAG context tokens ~67% while keeping a relevant passage.
Overview
A local MCP server that reranks and filters retrieved passages on your machine, cutting the context tokens sent to an LLM.
- What you can do
- Rerank, filter and compress retrieved passages against a query, keeping the top-K or a token budget and estimating the token savings before and after filtering. Count tokens exactly and compare or embed texts for near-duplicate detection. Rank a corpus against a query with lexical semantic search. Detect text language and classify text into supplied categories, with confidence signals.
- When to use it
- Useful when a RAG pipeline stuffs many retrieved chunks into every prompt and you want to send only the passages that matter. Best for single-hop questions where the query and answer share vocabulary; for multi-hop questions keep more passages rather than over-filtering.
- Requirements
- Runs locally as an npm package via npx on Node 18 or later; no GPU, build step or configuration. No network calls, API keys or environment variables are declared.
Installation
In SourceWeft
- Open AsterMind MCP in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
AsterMind MCP
Cut the tokens you send to an LLM — on your own machine, before the call goes out.
AsterMind MCP is a Model Context Protocol server that ranks and filters your retrieved context locally, so you send the model the few passages that actually answer the question instead of everything your retriever returned. It runs on the AsterMind Community Edition ELM engine. No network calls, no API keys, no data leaves the machine.
What it actually does (measured, not promised)
On a 6-scenario RAG benchmark (support KB, API docs, HR policy, e-commerce FAQ, DevOps runbook, fintech help), filtering 10 retrieved candidates down to the top 3 by rerank score:
Tokens counted with gpt-tokenizer
(o200k/cl100k BPE, OpenAI-compatible). Reproduce it yourself: npm run benchmark.
Full methodology and per-scenario results are in BENCHMARK.md.
The honest tradeoff
Savings and recall move in opposite directions — keep fewer chunks, save more tokens, risk dropping a relevant one. You choose the point on the curve:
Rank-1 surfaced a relevant passage in 100% of scenarios at every K we tested. That is the strong, defensible result. For single-hop questions (FAQ, support, one-fact lookups) keep top 1–3 and cut ~67–88% of context. For multi-hop questions that need several passages, keep top 5–6 — don't over-filter, or you'll drop a needed passage.
Install
Or add it to your MCP client config (Claude Desktop, Cursor, VS Code, Cline):
Runs on Node 18+. No GPU, no build step, no configuration.
Tools (10)
Every tool returns a confidence signal. Where the engine is weak, the tool says so rather than guessing silently.
Known limitations (stated on purpose)
- The reranker is lexical (TF-IDF). It matches on shared terms, so pure synonym gaps can be missed — a query for "payment methods" won't strongly rank a passage that only says "Visa/Mastercard." It excels when the query and the answer share vocabulary, which is the common RAG case.
- Classification is weak out of the box and only fair after training — use it as a low-confidence hint, not a decision-maker. It flags low confidence.
- Embeddings are character-level, useful for near-duplicate detection, not deep semantic similarity.
The token savings above are real only because a relevant passage is preserved. Cutting tokens by dropping the answer is not a saving, and this benchmark measures both.
When this saves you money
The server does not shrink an LLM's vocabulary or compress prompts magically. It saves tokens one specific way: you retrieve broadly, then send the model only the passages that matter. If your RAG pipeline currently stuffs 8–12 retrieved chunks into every prompt, reranking locally and sending 3 is a direct, repeatable cut to your per-call input tokens — and it runs on your machine for free.
Develop & verify
License
MIT © AsterMind AI. Built on @astermind/astermind-community.
Source: README.md at commit afda79b
Tools
0Version history
1- v0.1.0LatestSep 25, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts