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.
-
engramgithub.com/staticroostermedia-arch/engramupdated this year
Geometric agent memory for Agents: 8-tool lean contract, harness injection at wake, session handoff. 79 MCP tools. Not a vector DB. Rust + MCP.
AGPL-3.0 ★ 21 -
@michengai/dsh-engram-bundlegithub.com/michengai/dsh-engram-bundleupdated this year
NPM-installable DeepSeek Harness plugin that bundles Engram persistent memory MCP and injects parent-agent memory protocol.
npm Apache-2.0 -
LLM agent that investigates suspicious files over MCP and returns a cited verdict, with a prompt-injection red-team harness.
MIT ★ 1 -
DeepSeek Harness bundle that launches the compound-memory stdio MCP server, mounts the bundled compound-memory skill, and injects a short once-per-session usage reminder.
npm MIT -
opencode-memorygithub.com/xpajonx/opencode-memoryupdated this year
Distributed local-first hybrid-RAG memory layer and Harness bootstrapper for opencode agents, with dynamic pre-injection and styled console TUI.
★ 1 -
A prompt-injection firewall for AI agents, with no AI inside: your agent reads anything and takes orders only from you. For Claude Code, Codex CLI, Kimi Code, DeepSeek Harness, Gemini CLI, MCP hosts and LangChain.
npm MIT -
kakeibogithub.com/siash1/kakeiboupdated this year
A personal finance agent built from scratch on the raw Gemini API: hand-written tool-calling loop, double-entry PostgreSQL ledger, context caching, MCP server, 46-task eval harness, prompt-injection suite.
MIT -
Forward-deployed engineering in public: a role-scoped MCP server with prompt-injection defenses, 173 tests, and an eval harness (45.9% to 94.6% safe decisions). Live dashboard + threat model + ADRs.
-
dsh-memory-protocolactive
Memory protocol enforcement plugin for DeepSeek Harness — forces memory_weave before tool calls, auto-ingests session turns, injects memory context into agent steps. Bridges the opencode-memory MCP server with a hard protocol gate.
npm MIT -
dsh-mcp-authactive
Persistent MCP login state for DeepSeek Harness: OAuth 2.1 / manual header credentials per MCP server, injected into mcp-client rows via !!js. Settings UI talks to the host over Typert Remote.
npm MIT -
Project-scoped CodeGraph for DeepSeek Harness: launches codegraph's MCP server through npx, shares one instance per project root across sessions, selects the global codegraph version, and injects its tools only into workspaces that actually have an index
npm MIT -
MCP server + multi-agent orchestrator over FRED economic data: constrained tool contracts, a supervisor delegating to specialist agents, prompt-injection containment, a token-budget guardrail, and a CI-gated evaluation harness. Runs fully offline with no API key.
★ 1 -
Long-term memory plugin for DeepSeek Harness (MCP): L1 fact cards + L2 evidence timeline + L3 dialogue log + L4 episode summaries, 21 memory_* tools, per-turn auto weave/inject and scoped ingest, hybrid semantic+BM25+recency RRF retrieval, AES-256-GCM field encryption, local ONNX embeddings, zero API cost.
Apache-2.0 -
restaurant-agentgithub.com/illustraton916/restaurant-agentupdated this year
Permission-aware MCP server + spec-driven process + eval harness. Role-gated tools with structured refusals, prompt-injection evals against a real model — TypeScript
MIT -
Tool-calling SQL agent with read-only guardrails, an MCP server and an evaluation harness (bootstrap CIs, ablations, prompt-injection test)
-
Unofficial community plugin: per-session Engram memory for DeepSeek Harness — spawns an engram MCP child per agent session rooted at the session workspace, registers mem_* tools per agent scope, and injects the Memory Protocol as a system-prompt section.
MIT -
Run Claude Code on any LLM: a Cloudflare Workers AI gateway and LLM proxy that speaks the Anthropic Messages API and routes to Workers AI, OpenAI, Gemini, DeepSeek or Anthropic, with a prompt-injection session guard, red-team harness and MCP server.
MIT -
—
NOASSERTION ★ 2 -
Re-debugging Solved Failures? No Way. 1000+ Evidence-graded fix lessons, One MCP Call-->https://misakanet.org
npm Apache-2.0 ★ 561 -
hackbrowser-mcpgithub.com/badchars/hackbrowser-mcpupdated this year
The first browser MCP built for security testing. Give your AI agent a real Firefox browser and let it find vulnerabilities.
MIT ★ 24 -
Agent runtime guard — full-stack runtime interception for AI agents (Agent 运行时全栈拦截). A zero-dependency MCP server that blocks destructive commands, credential theft, SSRF and injection in real time, with Ed25519 evidence receipts. Works with Claude Code,
npm Elastic-2.0 -
Multi-agent vulnerability triage over MCP + A2A with a deterministic security harness; Docker/K8s deployable
MIT ★ 1 -
MCP 服务管理:增删改查与启停、连接测试、工具按需检索与全量注入、进程级与服务级环境变量(含密钥);MCP service management: CRUD, connection tests, tool injection, environment variables and secrets
npm MIT -
garbleworksgithub.com/SamsonCyber/garbleworksupdated this year
Authorized LLM red-team harness: recipes, scoped fire, evolutionary search, Wilson ASR. MCP + API + TUI.
Apache-2.0 -
Nothing reaches production without passing through the airlock. A change-control console where every irreversible production change is proven on a shadow copy in a sandbox before a human is ever asked to approve it. Built on TrueForge.
MIT ★ 1 -
Double-entry ledger with Kafka fraud detection and a local-LLM investigation agent (LangGraph + MCP), with an eval harness
MIT -
A stealth browser MCP for agents: three engines, diff observations, site memory, a human CAPTCHA handoff, and a benchmark harness.
oci MIT -
Exposes DSH as an MCP server: streaming thinking/tool/file notifications, shared intent, a floor lock that denies DSH writes while held, mid-turn alert injection, and review delegation.
npm MIT -
Capability-based credential broker and MCP server for coding agents. Use approved APIs without injecting secrets into prompts, argv, env, or agent code.
MIT -
Exposes DSH as an MCP server over Streamable HTTP: streaming thinking/tool/file notifications, shared intent, a floor lock that physically denies DSH writes while held, mid-turn alert injection, and review delegation.
npm MIT -
Local-first, file-native long-term memory hub: preflight, compact context injection, citations, human-approved commits, temporal validity, dedupe/merge, retrospective reflection, write-risk flags, a derived SQLite index, per-workspace scheduled distillati
npm MIT -
MCPGoatgithub.com/SabyasachiDhal/MCPGoatupdated this year
Deliberately vulnerable MCP server (aka MCP Goat) for security training — 26 challenges across 4 difficulty levels (incl. a secure reference), a victim-agent harness, and one-command Docker deploy. Practice penetration testing against the Model Context Protocol.
MIT ★ 12 -
Codebase Q&A with (path:line) citations — the capstone of a from-scratch AI-engineering series. Built in 8 eval-gated stages (chat → RAG → agents → hardening → production), provider-agnostic and offline-first, with 4 extensions (MCP, agent harness, context management, local models).
MIT ★ 1 -
Self-improving context & memory for Claude Code and OpenAI Codex. One canonical knowledge store injected only when relevant — hooks push budgeted context, an MCP server serves depth on demand, and it learns lessons from your sessions. Agent memory + context engineering as a single static Go binary.
MIT ★ 3 -
Claude agent that finds, fixes and verifies faults in a simulated 100-vehicle fleet through a Go MCP server. Code-enforced write and stale-data checks, a ground-truth eval harness, and a rule-based baseline: Claude 6/6 vs rules 0/6 when the fix is only in the manual.
Apache-2.0