Raven

At a glance

Runs where

On their server. What you send to it goes to them.

What you need

Nothing: it answered without a login

Tools

45 tools

Its own safety labels

45 read-only

Memory it takes

About 16,300 tokens of your AI's memory, just for its tool descriptions

Largest answer

About 52.1 KB in our test

Speed

Average answer in our test: 225 ms

Tokens are an estimate: the characters in its tool names, descriptions and inputs, divided by 4. The exact number depends on your AI model. Safety labels are what the server says about itself. From our latest test, 5 Nov 2026.

Is Raven MCP server safe?

• None of the hidden-instruction wordings we know were found in its 45 tool descriptions.
• Its error messages showed no secrets or internal details.
• All 45 of its tools are labeled read-only.
• Overall it scored 83/100 on 5 Nov 2026.

A test can only see what the server shows from outside, not what it does with your data behind the scenes. Connect servers from makers you trust.

What this test does not check

• Its source code and packages.
• What a tool quietly does behind the scenes, such as writing a file or calling another service.
• Instructions hidden inside live tool answers. We only look for those in tool descriptions.
• Hidden instructions worded in a way our list doesn't know yet.
• Tools that could change something. Those are read, never run.

What QAmcp can and can't see · How to tell if an MCP server is safe · Hidden instructions · Prompt injection

Known security advisories

None found in GitHub Security Advisories and OSV.dev on 5 Nov 2026.

That means none were published there, not that there are no problems. QAmcp tests how a server behaves. It does not scan source code.

What it can do

45 tools, in the server's own words.

and 21 more

Test results

Safe mode: only tools labeled read-only were run. How the test works

• !
  Ready for the newest MCP version

  1 warning

  • !Uses an older MCP version (2025-11-25). It works today, but it isn't on the July 2026 version (2026-07-28) yet.
  note
• !
  Every tool is described

  4 warnings

  • !audit_page: Inputs with no description: viewport
  • !audit_parity: Inputs with no description: ios, android, checklist
  • !audit_ios_a11y: Inputs with no description: elements, viewport, options
  • !audit_screen: Inputs with no description: scroll_settle
  note
• !
  Fast enough

  1 warning

  • !audit_responsive_visibility: Slowest tool: audit_responsive_visibility (4.6 s): users will notice the wait
  • and 3 more
  note
• !
  Follows MCP best practices (not scored yet)

  2 tips (not scored)

  • !The tool list is about 67 KB (roughly 17k tokens). AI apps send it with every message, so it takes that much of the AI's memory before any work starts.
  • !25 tools answer with structured data but don't declare its shape (outputSchema): get_principles, get_pattern, search_knowledge, get_checklist, get_d4d_framework, list_design_systems and 19 more. If those answers change shape, nothing can ca
  note
• !
  Resources and prompts work

  3 warnings (5 prompts)

  • !cadence: Never uses the input project: what you type there doesn't appear in the result.
  • !plumage: Never uses the input project: what you type there doesn't appear in the result.
  • !truesight: Ignores all its inputs (profile, project): what you type never appears in the result. Check the input is actually used.
  • and 1 more
  note
• ✓
  ConnectsConnected in 540 ms
  pass
• ✓
  Lists its toolsFound 45 tools
  pass
• ✓
  Tools are labeled safe or dangerousEvery tool has a safety label
  pass
• ✓
  No known hidden instructionsNone of the hidden-instruction wordings we know were found in tool descriptions. New wordings can get past this check.
  pass
• ✓
  Tools work with good inputs

  All good

  • ✓get_principles: Worked (118 ms)
  • ✓get_pattern: Worked (123 ms)
  • ✓get_business_strategy: Answered with an error, probably because of our sample input: { "error": "Strategy type 'test' not found.", "available_types": [ "growth", "metrics", "monetization", "onboarding", "retention" ] }
  • ✓evaluate_design: Answered with an error, probably because of our sample input: { "error": "nothing to evaluate — no description and no before/after screenshot pair. Pass a description of the design, or pass before_scre…
  • ✓search_knowledge: Worked (126 ms)
  • ✓get_checklist: Worked (128 ms)
  • ✓get_d4d_framework: Worked (115 ms)
  • ✓list_design_systems: Worked (149 ms)
  • and 32 more
  pass
• ✓
  Bad inputs give a clean error

  All good

  • ✓get_principles: Refused missing required input "context" with a clear error
  • ✓get_principles: Refused wrong type for "context" with a clear error
  • ✓get_pattern: Refused missing required input "type" with a clear error
  • ✓get_pattern: Refused wrong type for "type" with a clear error
  • ✓get_business_strategy: Refused missing required input "type" with a clear error
  • ✓get_business_strategy: Refused wrong type for "type" with a clear error
  • ✓evaluate_design: Refused wrong type for "description" with a clear error
  • ✓search_knowledge: Refused missing required input "query" with a clear error
  • and 32 more
  pass
• ✓
  Tools work together

  3 chains working

  • ✓list_design_systems → get_design_system: get_design_system found what list_design_systems returned (id = "stripe")
  • ✓list_design_systems → audit_responsive_visibility: audit_responsive_visibility found what list_design_systems returned (url = "https://stripe.com")
  • ✓list_design_systems → audit_contrast: audit_contrast found what list_design_systems returned (url = "https://stripe.com")
  • ✓Stopped looking for more chains after the limit on tool calls
  pass
• ✓
  No secrets leakedChecked 111 error messages: no secrets or internal details
  pass

Test it yourself

Free and open source. Same checks, same score.

Show this score on GitHub

Free for every server in the directory. Paste this into your README: the badge shows the score from our latest test.

Score over time

• 5 Nov 202683

One test so far. New tests will show here.

Reviews

Coming soon: reviews from people who use Raven, signed in with GitHub, one per person. Reviews never change the score.