MCP servers that ship as code — a repository or a package you install and run locally (usually over stdio). They have no hosted endpoint, so we cannot probe or grade them: instead each one shows its repository, install command, license, stars and when it was last updated.
Collected from GitHub repositories and npm, PyPI, Cargo, Go, Maven, NuGet and RubyGems packages tagged as MCP servers, the official MCP Registry, Docker's MCP catalog, ModelScope, Baidu MCP World, Cline and other public directories; one entry per repository. A server that also offers a hosted endpoint is listed under Remote instead. These entries are not counted in our reachability, verification or readiness statistics.
-
Autonomous Agentic Operating System — agents as processes with cost control, governance, observability, MCP, SDK, CLI, and skills. Fail-closed. Verified end-to-end.
NOASSERTION ★ 1 -
sqlgategithub.com/asd3358123/sqlgateupdated this year
A deterministic, fail-closed write gate for LLM-generated SQL, exposed over MCP. The model can't talk its way past it.
MIT -
Auditable deterministic reasoning MCP server (non-LLM). World-graph → A* auditable reasoning → execution → learning loop. Fail-closed: an unchecked claim is never returned as passed. Free for non-commercial agent use.
NOASSERTION -
Unofficial local-first iFood MCP. Checkout fail-closed.
MIT ★ 1 -
AdversaryGate as an MCP server: a fail-closed merge gate an agent calls before saying done
pypi AGPL-3.0-only -
Local MCP server for verifiable DOCX redline history and fail-closed tracked-change edits
pypi Apache-2.0 -
Install LBrain where the agent already lives. Native MCP, companion skill, fail-closed session start.
BSD-3-Clause -
IQA attestation primitives (derived intent addressing + AE-128 envelope verification) as Model Context Protocol tools. Offline, fail-closed, stateless.
pypi Apache-2.0 -
Fail-closed image updates for Portainer and Compose. A digest ripens. You apply.
MIT -
Open-source MCP server and CLI for ManyChat, with a fail-closed Meta messaging-policy guard.
AGPL-3.0 ★ 1 -
Self-hosted MCP proxy for Basic Memory with explicit, fail-closed context routing and project lifecycle management.
MIT -
IQA attestation primitives (derived intent addressing + AE-128 envelope verification) as MCP tools. Offline, fail-closed, stateless.
Apache-2.0 -
mcp-databasegithub.com/DiegoBulhoes/mcp-databaseupdated this year
MCP server for PostgreSQL, MySQL and MongoDB with access modes, fail-closed query validation and audit logging
Apache-2.0 -
truth-social-mcpgithub.com/devrim-1283/truth-social-mcpupdated this year
Policy-aware MCP 2.x server for authorized Truth Social access — read-only by default, fails closed without explicit authorization.
MIT ★ 1 -
🛡️ Ephemeral privacy & security middleware for Model Context Protocol (MCP). Zero-persistence credential masking & fail-closed execution guard for AI agents.
MIT -
umbryn-mcpgithub.com/rinava/umbryn-mcpupdated this year
Redacts PII/PHI from text before it reaches an LLM. Fail-closed, self-hosted, no egress.
pypi MIT -
evidra-lockgithub.com/vitas/evidra-lockupdated this year
MCP Kill-switch for AI agents. Validates infrastructure operations before execution. Fail-closed. Evidence-backed.
Apache-2.0 ★ 8 -
governed-contextgithub.com/leonardoeverling-spec/governed-contextupdated this year
Governed memory for AI agents: explicit states, fail-closed sensitivity gates, quarantine-only writes (dependency-free MCP server) + a blind-evaluation protocol skill
MIT -
Exact-decimal double-entry kernel for AI agents. Call MCP tools instead of doing money math in tokens. Fail-closed, audit-hashed, deterministic.
MIT ★ 1 -
umbryn-mcpgithub.com/Rinava/umbryn-mcpupdated this year
MCP server for redacting PII/PHI and de-identifying healthcare text before it reaches an LLM. self-hosted, fail-closed, HIPAA-aware.
MIT -
phi-redact-mcpgithub.com/rinava/phi-mcpupdated this year
Fail-closed MCP server that redacts PII/PHI from text before it reaches an LLM — self-hosted, no egress, HIPAA-aware.
pypi MIT -
Minimal, secure HTTP & SOCKS5 proxy routed through Tor with fail-closed leak protection, non-root execution, and a REST API for circuit rotation and dynamic exit node selection.
MIT ★ 7 -
Security-hardened MCP server for terminal access: deny-by-default policy engine, shell=False sandboxed execution, cwd jail, secret redaction, fail-closed audit log.
MIT -
MCP gateway for Codex to orchestrate external coding-agent harnesses, with normalized lifecycle, localhost settings UI, and fail-closed no-overage safety.
MIT -
Multi-node control plane and MCP gateway for DSH ChatGPT Helm connectors: hub + node agents, authenticated WS JSON-RPC, compact session summaries, queue/steer routing, fail-closed routing
★ 1 -
Safe, screenshot-grounded computer-use MCP server that lets AI agents see, decide and act on the Windows desktop — with semantic verification, fail-closed safety gates and approval controls.
MIT ★ 4 -
Unified Korean stock broker SDK (KIS, DB Securities, Toss...) with fail-closed safety primitives and a read-only-by-default MCP server. Apache-2.0.
Apache-2.0 -
Policy hook (fail-closed for checked tools) and tamper-evident evidence log for Claude Code. The code runs locally. Evidence supporting audits. Not a sandbox. PolyForm Noncommercial.
NOASSERTION -
@eternal-roman/ledger-mcpgithub.com/eternal-roman/ledgerupdated this year
MCP server for the Ledger kernel. Agents call money_compute, entry_validate, ledger_post — they do not do money math in tokens. Fail-closed, audit-hashed, deterministic.
npm MIT -
Fail-closed human approval gates for AI agent actions, as an MCP server. Default-deny, TTL expiry, append-only audit log.
npm MIT -
Vietnam stock market MCP server (24hMoney, 13 tools) with fail-closed API-key auth — one image deploys to AgentBase Agent Runtime, vServer/VKS in your VPC, or on-premises.
MIT -
sideguardgithub.com/alisaitteke/sideguardupdated this year
MCP guard & vibe coding security tool for Cursor and Claude Code. Fail-closed human-in-the-loop approval for shell commands and MCP tools—local daemon, YAML policy, no cloud.
MIT ★ 2 -
Governed MCP server for IP cameras (VAPIX + ONVIF SOAP): fail-closed per-agent policy, hash-chained signed receipts on every call — including denials. First reference producer for the AAR spec.
MIT ★ 4 -
Local-first memory for DeepSeek Harness: per-turn BM25 recall you can audit — open-rate ledger, memory report, fail-closed write guard, 14-check doctor. Zero deps, no network, no model calls unless you opt in.
MIT -
Local-first, human-approves-everything grocery replenishment as a ResonantOS add-on: runway analysis, fail-closed dietary checks, and a next-cart draft that structurally cannot check out. No accounts, no cloud, no telemetry — nothing leaves your machine.
MIT -
Correlates every NVD CVE with CWE, CAPEC, ATT&CK, D3FEND, OWASP, CISA KEV, Vulnrichment SSVC, and EPSS, with APT links only where MITRE ATT&CK cites the CVE. Search-first web app, MCP server, daily fail-closed updates via GitHub Actions.
NOASSERTION ★ 1 -
Local, zero-token event triggers for Claude Code, Codex, OpenCode and ZCode: when CI fails, a training run finishes, a service goes down or NVDA crosses your price, your agent wakes up in your conversation—even if closed.
MIT ★ 8 -
Approvals, spend limits and a kill switch for AI agents, set up in one command: npx immiscible init. See it first, offline and with no account: npx immiscible try. Installs a fail-closed Claude Code hook, prints MCP setup for Claude Code, Cursor and VS Co
npm MIT -
MIT-licensed MCP bridge to the Alethia runtime — the patent-pending zero-IPC E2E test runtime for AI agents. 2-5x faster than Playwright MCP. Signed evidence packs, EA1 fail-closed safety gate, WCAG + NIST 800-53 audits built in. Local-first, zero telemet
npm MIT -
AI advisory board as a Claude Code skill + MCP server: verifies every quote word-for-word against public-domain sources, or abstains. Fail-closed fidelity. | Совет советников как скилл для Claude Code и MCP-сервер: каждая цитата сверяется дословно с источником в общественном достоянии, иначе отказ.
MIT ★ 10 -
Open-source MCP server for Adobe Experience Platform, Journey Optimizer and Customer Journey Analytics — 61 tools across 14 categories, full read AND write, from one OAuth credential. Three write modes with a fail-closed sandbox guard that will not write to production by default. Ships AEC-Bench, an agentic benchmark. Apache 2.0.
Apache-2.0 -
Verax: the accountable agent body. Gate, decision record, receipt, reconciliation. Apache-2.0.
npm Apache-2.0 -
Executed-evidence verification gate for AI coding agents
AGPL-3.0 ★ 4 -
Local admission gate for software agents. Rust, HTTP and stdio MCP integrations with policy, operator revocation and structured evidence.
Apache-2.0 ★ 5 -
halugardgithub.com/ppp727276/halugardupdated this year
Anti-hallucination MCP server for AI agents. 4-gate dual-stage pipeline. Go binary. Blocks fabricated claims, identity drift, credential leaks, dangerous commands.
MIT -
Alpha MCP server for local-first LLM orchestration: memory, knowledge, state, routing, swarm planning, API discovery, tests, automation, and plugins.
npm MIT ★ 2