MCP servers · shortlist
MCP servers for databases
Database connectors, data platforms, filesystem bridges and the knowledge stores an agent reads from. These are the servers that decide whether your agent can see your actual data.
How this list was ordered. There is no reliable popularity signal in the source registries, so entries are ranked by how completely the record is documented — a substantial description plus a documentation, repository or base URL — and then alphabetically. 464 mcp servers in this scope clear the content bar; the top 24 are shown. A place here is not an endorsement or a paid slot.
Shortlisted entries
- 01
juyterman1000/entroly
Auditable context control plane and MCP server for AI coding agents. Compresses context 70–95% (BM25 + entropy + dep-graph knapsack), stabilizes prompt prefixes for provider cache discounts, routes easy tasks to cheaper models (RAVS Bayesian router), and verifies answers locally with WITNESS hallucination guard (0.844 AUROC, $0, ~3 ms). MemoryOS adds local budget-aware working/episodic/semantic memory with decay, safety scanning, and durable persistence. 38 agent integrations (Cursor, Claude Code, Codex, Aider, and more). Ships as MCP server (entroly serve), HTTP proxy, or Python/Rust library. Apache-2.0, local-first. pip install entroly
Knowledge & Memory auth: None
- 02
Eszetael/postgres-mcp-hardened
Maintained Rust replacement for the archived @modelcontextprotocol/server-postgres. Writes are refused twice: sqlparser AST validation rejects mutating statements before execution, and the session runs defaulttransactionreadonly with a per-session statementtimeout — so the read-only transaction is the fallback, not the only defence. Single self-contained binary, stdio and Streamable HTTP, schema inspection, column redaction that survives renames and casts, an EXPLAIN cost guard, a hash-chained audit log with --verify-audit, optional OAuth 2.1 RS256, distroless image, signed releases for five platforms. MIT.
Databases auth: OAuth
- 03
vornicx/Midas
Local-first memory for long-horizon AI agents with no LLM at ingest — $0 per message, zero data egress, verbatim source-traceable recall. Typed belief revision, selective forgetting, and a provenance guard that blocks memory-justified destructive actions unless user-confirmed. One midas init wires Claude Code, Cursor, Codex, and 6 more clients to one shared SQLite memory; local web inspector and hash-chained audit log. Benchmarked on LongMemEval/LoCoMo/BEAM with one-command repros — including published negatives. uvx midas-memory-mcp or npx -y midas-memory-mcp
Knowledge & Memory auth: None
- 04
Vasilispapg/claude-handoff
Post-hoc session handoff and project memory for Claude Code: turns any session transcript — including crashed or usage-limit-hit ones — into one clean handoff.md another model continues from (ChatGPT, Gemini, or the next Claude), and distills a project's ENTIRE session history into one cited memory brief injected at every session start. MCP tools listsessions and handoff (LLM summaries only behind an explicit --allow-llm); zero dependencies, pure stdlib, secrets redacted from every output. pipx install claude-handoff, then claude-handoff --mcp.
Knowledge & Memory auth: None
- 05
ni-c/wikijs-mcp
Search, read and edit a self-hosted Wiki.js 2.x, plus its assets, comments, users and groups (62 tools). Wiki.js' default search engine indexes only titles and descriptions, so searchpages names the active engine and greppages searches what is actually written in the pages. updatepage compares against the moment you read a page and refuses to overwrite an edit saved in between; surgical find-and-replace must match exactly once. Read-only mode, a write path scope, and confirmation tokens on everything destructive. npx -y @ni-c/wikijs-mcp
Knowledge & Memory auth: API Key
- 06
Battam1111/Myco
Agent-first cognitive substrate with 18 manifest-driven verbs (germinate / eat / assimilate / sporulate / traverse / immune / molt / …) and 25 lint dimensions enforcing contract invariants mechanically (R1–R7). Cross-session / cross-project memory via a self-validating filesystem graph — AST + markdown-link derived, not embedding-based. Provider-agnostic by design: MP1/MP2 dims forbid LLM-SDK imports in the kernel and plugin tree. Editable-default install. Works with Claude Code, Cursor, Windsurf, Zed, VS Code, and any MCP client.
Knowledge & Memory auth: None
- 07
vishalbanwari26/mnemos
Persistent memory with three swappable storage backends (Postgres/pgvector, Qdrant, Neo4j) behind one interface, benchmarked head-to-head for latency, not just claimed interchangeable. Retrieval weighting adapts per user via epsilon-greedy over real outcome feedback instead of a fixed formula. Reflection merges near-duplicate facts and decays/archives stale ones, audit-logged, never hard-deleted. Real before/after MCP eval: +60% answer accuracy with memory vs. without. git clone + uv sync (not yet on PyPI).
Knowledge & Memory auth: None
- 08
FROWNINGdev/django-orm-lens
Django ORM static-analysis MCP server. Nine read-only tools (listapps, listmodels, describemodel, findrelations, cascadepreview, erdiagram, describemigrationdependency, suggestindexes, signalgraph) that expose Django schema, relationships (FK/M2M/O2O, ondelete, relatedname), migration deps, missing indexes, and signal graphs — no DB, no Django boot, no credentials. Companion to a VS Code / VSCodium / Cursor extension (Open VSX verified) and a django-orm-lens PyPI CLI. pip install django-orm-lens[mcp].
Databases auth: None
- 09
vishnu-77/openreflex
Local execution memory for AI coding agents (Claude Code, Codex CLI, Cursor, OpenCode). Lifecycle hooks record coarse task metadata into a SQLite Experience Graph (no code, commands, tool output or transcripts, and no model calls). Before a similar task it injects what worked, the relevant files and known fixes with an execution budget; during a task it recommends whether to continue, pivot or stop. MCP tools cover execution context, progress checks, outcomes, search and explanations.
Knowledge & Memory auth: None
- 10
oomkapwn/enquire-mcp
Long-term memory for AI agents (Claude Code/Desktop, Cursor, ChatGPT, Codex, OpenClaw) backed by a local Obsidian markdown vault. Hybrid retrieval (BM25 + ML embeddings + BGE reranker, RRF-fused), HNSW + int8 quantization, agentic RAG (HyDE + sub-question), GraphRAG-light (Louvain wikilink community detection), standalone Obsidian Bases, PDFs + Tesseract OCR. 46 tools, 19 MCP prompts, MIT, SLSA L2, zero cloud calls during serve. npx -y @oomkapwn/enquire-mcp serve --vault <path>
Knowledge & Memory auth: None
- 11
TT-Wang/cortex-plugin
Persistent, self-evolving memory plugin for Claude Code. Background miner extracts durable lessons (decisions, conventions, bug fixes) from completed sessions via Claude Haiku, stores them as human-readable markdown in an Obsidian vault, and assembles query-tailored context briefings at session start. Local-first, no cloud, no API keys. Self-healing install via uv bootstrap shim, /cortex-doctor preflight, graceful FTS-only degraded mode when claude CLI missing. MIT.
Knowledge & Memory auth: API Key
- 12
dahshanlabs/klypix-mcp
Shared, versioned project brain for coding agents: one brain.klypix file committed with the repo holds current decisions, corrections that supersede (archived, never deleted), evidence anchors that flag drift when cited code changes, and open questions. Sessions declare their task and files and are warned about exact-file overlap. Lifecycle hooks for Claude Code; MCP config + rules for Codex, Cursor, Gemini CLI, Cline and Copilot. npx klypix-mcp install
Knowledge & Memory auth: None
- 13
Evan-Crx/permisapi-mcp
7 tools for French open-data building permits (Sitadel, 311k rows, ~2M permits/year, Etalab license). Search, details, DVF transactions cross-ref, real estate dealer opportunity score, PLU urban zoning, BRGM natural and technological risks, and a Vue 360 composite that fans out 6 sub-fetches in one tool call. Powered by permisapi.fr. Plan Free covers basic search and details, Pro+ unlocks the enrichments. Install: pip install permisapi-mcp.
Data Platforms auth: None
- 14
SVerITG/Metis
A private, local research "second brain" for Claude: project-aware memory, cited answers from your own library (won't invent what it can't find), linked notes/meetings/ideas via a domain-specific knowledge layer, daily briefs (news + new papers in your field), a live meeting assistant, cross-pollination across your work, and 34 routed agents. A governed layer between you and the AI, with guardrails like data protection. A Research Cortex.
Knowledge & Memory auth: None
- 15
AkashGoenka/coldstart
Codebase memory for coding agents, with no embeddings and no API key. A deterministic AST index answers "which files are relevant to this task?" in milliseconds, and agents write durable notes about the repo that are content-hash checked — a note flags itself stale the moment the code it describes changes. Notes are markdown inside the repo, so they commit and review alongside your code. npm i -g @cstart/coldstart && coldstart init
Knowledge & Memory auth: API Key
- 16
CodeAbra/iai-personal-memory-engine
Local memory daemon for any MCP-over-stdio client with three-tier storage (episodic/semantic/procedural). Own SQLite + hnswlib store (Hippo) with bge-small-en-v1.5, MIT-licensed community-detection reranking (MOSAIC), and sleep-cycle consolidation. AES-256-GCM encrypted at rest, no telemetry. Verbatim recall >=99% and post-contradiction Rescue@10 1.000 at honest scale. Ambient capture via shell hooks. Windows support in beta.
Knowledge & Memory auth: None
- 17
dbwls99706/deadends.dev
Structured knowledge of what does NOT work. 2,400+ entries across 54 domains record the dead ends for an error - the fix that looks right and fails, and why - alongside workarounds with success rates and an error transition graph. Covers code errors and country-specific real-world rules (visa, banking, legal, emergency) where generic answers are most often wrong. pip install deadends-dev, or hosted at https://deadends.dev/mcp
Knowledge & Memory auth: None
- 18
DomDemetz/claude-soul
Self-improving learning engine for Claude Code. Extracts signals from every session (corrections, successes, confusion), runs periodic reflections, and evolves behavioral frameworks through evidence tiers (hypothesis → observed → validated). Frameworks that keep working get promoted, bad ones get retired. 9 MCP tools, automatic hooks, phase-adaptive learning. Single dependency, local-only. npx claude-soul init --starter
Knowledge & Memory auth: None
- 19
wenb1n-dev/SmartDBMCP
A universal database MCP server supporting simultaneous connections to multiple databases. It provides tools for database operations, health analysis, SQL optimization, and more. Compatible with mainstream databases including MySQL, PostgreSQL, SQL Server, MariaDB, Dameng, and Oracle. Supports Streamable HTTP, SSE, and STDIO; integrates OAuth 2.0; and is designed for easy customization and extension by developers.
Databases auth: OAuth
- 20
DanceNitra/inspeximus
Persistent agent memory with a first-class correction channel: supersede or revert a fact so recall stops returning the stale value, an echo guard that blocks a restated old value from resurrecting it, and receipted erasure (a signed, content-free tombstone that makes a deletion provable). No LLM on the write path, zero dependencies, the core in one file, 73 MCP tools. uvx --from "inspeximus[mcp]" inspeximus-mcp
Knowledge & Memory auth: None
- 21
Mnemosyne-OS/Mnemosyne-Neural-OS
Memory the human governs: vaults separated by domain (code, notes, journal) on your own machine, hybrid retrieval (BM25 fused with vectors by RRF), and a declared vault list, so a vault left out cannot be read by mistake. Also reads what the OTHER coding agents on this machine wrote to disk, and flags when two are live in the same git worktree. Needs the Mnemosyne OS desktop app running. npx -y @mnemosyneos/mcp
Knowledge & Memory auth: None
- 22
Kemetra/seshat-bi
Read-only readiness governance for BI pipelines. Six tools report where each table sits across a seven-stage source-to-Power-BI spine, explain what is blocking the next stage, run a static SQL/TMDL/PBIR governance check, and export an evidence pack. Tools never write files, execute warehouse work, or grant approvals -- human sign-off stays a separate seam. Install: pip install "seshat-bi[mcp]" then seshat mcp.
Data Platforms auth: None
- 23
gingugu/gingugu
Persistent memory for AI coding assistants. Local SQLite, no cloud. 16 MCP tools: store, recall, search, relate, consolidate, export, and credential vault (OS keychain). Typed memories with confidence lifecycle (verified/inferred/stale/deprecated), namespaces, knowledge graph, and hybrid BM25 + semantic search via fastembed ONNX. Works with Cursor, Windsurf, Claude, and any MCP client. pip install gingugu
Knowledge & Memory auth: None
- 24
gulmezeren2-byte/erp-report-engine
Read-only MCP over the SQL database behind an ERP (Logo Tiger, Netsis, Mikro). The agent sees canonical entities like orders, never raw ERP tables, through a four-layer read-only guard that checks the statement, fails closed, and blocks side-effecting functions (pgreadfile, xpcmdshell, …) — measured by a public 28-attack benchmark and an in-browser "break it" playground running the real guard via Pyodide.
Databases auth: None
Snapshot 2026-09-18. Fields shown are what the source registry stated, not a live probe — check the linked documentation before integrating. See the methodology.