Perseus Computing releases local-first MCP servers

Mimir is an 8 MB Rust binary (v2.7.0) that encrypts all stored facts with AES-256-GCM and bundles a dense-embedding model directly inside the executable [Mimir GitHub]. It runs without an API key, model download, or network connection, and uses a SQLite + FTS5 hybrid for bi-temporal search. The FTS5 index is built over plaintext, while the underlying rows remain encrypted at rest—a trade-off the authors call out in the docs [DevTo]. Perseus is a pip-installable Python package that compiles repository state, service descriptors, and test results into a deterministic briefing at session start [Perseus GitHub]. The compiled context is byte-stable: identical repo snapshots always produce the same context payload. Both tools are listed on the official MCP Registry and can be paired with any MCP client, including Claude Code, Cursor, and Cline. Benchmarking on offline harnesses shows Mimir’s bundled embeddings achieve 91.7 % recall@1 and 100 % recall@5 on paraphrased queries, while a naive keyword search on the same data reaches only 4.2 % recall@1. Perseus maintains full answer coverage at a fixed context size where tuned RAG baselines begin to drop facts. By keeping memory and context on-prem, Mimir and Perseus eliminate the barrier to regulated firms shipping data to external clouds without violating compliance regimes [DevTo]. Deterministic context compilation removes the token-burn of repeated tool-call discovery, reducing latency and making performance predictable for latency-sensitive workloads. The bi-temporal model lets teams query “as-of” any point in history, enabling audit trails without manual versioning [Mimir GitHub].