← Back to directory
COMPARE UP TO 4 SERVERS

Compare MCP servers

Compare scores, permissions, risks, and fit in one decision-focused table.

DimensionMARM Memory MCP ServerLocal-first persistent memory and semantic recall for AI agentsContext7Upstash's official server providing up-to-date third-party library docs for AI coding assistantsNPM Sentinel MCP ServerAI-powered NPM package analysis MCP server
FMRS63 / 100 · C80 / 100 · B79 / 100 · B
Reliability10 / 2014 / 2012 / 20
Security and permissions11 / 2016 / 2016 / 20
Maintenance13 / 2017 / 2018 / 20
Documentation15 / 2015 / 2018 / 20
Setup experience14 / 2018 / 2015 / 20
Best for
  • Developers who want memory to stay entirely local instead of depending on a cloud vector database
  • Users who switch between multiple MCP clients and need context to carry over
  • Agent users who need code-structure awareness and change-impact analysis
  • Self-hosters comfortable running a Python package or Docker container
  • Developers using fast-moving frameworks/libraries worried about the AI suggesting stale code
  • Scenarios wanting zero-config documentation lookup
  • Developers auditing NPM dependencies within AI workflows
  • Teams performing supply chain security assessments
  • Users of Claude Desktop, Cursor, or VS Code
Not for
  • Teams that require an officially hosted cloud service or SLA
  • Pure API integrations that do not use an MCP client
  • Restricted environments that cannot run local processes or containers
  • Users expecting a fully managed zero-configuration cloud product
  • Looking up internal/private codebase documentation (Context7 targets publicly published open-source libraries)
  • Cases needing very high coverage of obscure, niche libraries (coverage depends on what Context7's platform has indexed)
  • Projects outside the NPM ecosystem (e.g., pure Python/Go)
  • Environments without network access to deps.dev, OSV.dev, and the npm registry
  • Scenarios requiring maintenance by an official upstream vendor
Required permissions
  • Read/write access to the local data directory (default ~/.marm/, or %USERPROFILE%\.marm\ on Windows) holding the SQLite memory database, index database, and logs
  • HTTP mode binds to localhost:8001 by default and can be exposed to the network with SERVER_HOST=0.0.0.0
  • Docker deployments need a mounted data volume, and code indexing needs read-only repo mounts
  • Network access on first code-graph use to download the graph engine binary (~269MB, one time)
  • Docker HTTP mode requires a MARM_API_KEY environment variable for Bearer authentication
  • Usable without an API key (subject to a free-tier rate limit); CONTEXT7_API_KEY is an optional credential for higher quota
  • Read-only documentation lookup — no code execution or local filesystem access involved
  • Network access to the NPM registry and external services (deps.dev, OSV.dev, OpenSSF, npms.io, GitHub)
  • Read access to workspace lockfiles (pnpm-lock.yaml, package-lock., yarn.lock) for cache invalidation
Risks and side effects
  • Exposing the HTTP port with SERVER_HOST=0.0.0.0 without a firewall and TLS proxy can leave memory data reachable by unauthorized parties
  • Memories and code index data are stored as plaintext SQLite files that any process able to read the directory can access
  • In Docker, code-graph tools can only see mounted container paths; using host paths causes index failures
  • Enabling write-time consolidation (CONSOLIDATION_ENABLED=1) raises median write cost from about 6.5ms to 58.1ms
  • Recall latency temporarily increases while the concept-graph backlog drains (measured about 8ms to 16ms median on a real corpus)
  • Graph tools degrade with an error if the graph engine fails to start, for example with no network on first download, a full disk, or schema drift
  • The free tier has limited quota — high-frequency use may hit rate limits
  • Documentation content comes from Context7's platform index, so its accuracy and freshness depend on that platform's crawl cadence
  • Third-party READMEs and changelogs are untrusted external content; they are wrapped in tags with _meta flags but should still be handled cautiously
  • Depends on availability and accuracy of external services
  • Batch requests are capped at 25 packages to prevent registry enumeration; larger sets require batching
  • Third-party open source project, not officially maintained by NPM or Anthropic
Supported clientsClaude Code, VS Code / GitHub Copilot Agent, Cursor, Codex CLI, Gemini CLI, Qwen Code, xAI / Grok Remote MCPClaude Code, VS Code, Cursor, Cline, AmpClaude Desktop, VS Code, Cursor, Smithery.ai
Tools14219