Delphi.
An open-source MCP context engine.
Index code, docs, papers, and datasets. Serve your agents the right context, ranked. Built for the kind of research where retrieval is half the answer.
One index. Four sources.
Most retrieval engines specialise. They are good at code, or good at docs, or good at semantic search over PDFs. Delphi is built to hold all four in the same index. Code, documentation, papers, and datasets, so an agent can answer a single question by reaching across them.
It speaks Model Context Protocol and plain HTTP. Plug it into Claude Code, Cursor, Windsurf, or any agent that grounds its answers in source material.
Numbers.
We maintain a small open benchmark, the SynsciContextBench. Eleven phases (code retrieval, multi-hop, SWE-Agent code generation, diff-aware re-indexing, real-session replay, and more) against a fixed corpus of two popular Python libraries and their documentation. Same index, same embeddings, same queries. The numbers below are from the last full run.
Reproduce: git clone github.com/synthetic-sciences/SynsciContextBench && make bench.
Inside the engine.
Delphi is FastAPI in front of Postgres with pgvector for embeddings and tree-sitter for symbol extraction. The retrieval side is a hybrid: dense vectors, BM25 over text, trigram for fuzzy identifiers, and an exact-symbol channel. Results are fused per branch, then re-ranked by a cross-encoder.
Documents are chunked by heading where headings exist, by symbol where they don't. The chunker preserves enclosing class and linked tests, so a hit comes back with the context an agent actually needs to act on it.
For longer work, Delphi exposes an async research surface over server-sent events. Agents POST a question, get a session id, then stream iteration events as the engine searches, reads, and decides.
Quickstart.
One command. No accounts, no API keys, no cloud. The installer writes the MCP config for every agent it finds on your machine and brings up the local index.
# install + run locally$ npx @synsci/delphi # index a repository, auto-discovering its docs$ delphi index github.com/fastapi/fastapi --auto-docs # the MCP config for Claude Code / Cursor / Windsurf# was written automatically. restart your agent.
Prefer Docker or the Python client? See the install guide.
Apache 2.0.
Delphi runs on your laptop, in your VPC, or anywhere Docker runs. Your index lives on your hardware. No telemetry, no auth wall, no rate limit.
github.com/synthetic-sciences/delphi. Issues, pull requests, and forks all welcome.
