ProtocolSutra AI
Check captured MCP and A2A files for contract breaks and hostile tool text.
Offline lab for captured MCP and A2A files. npm [email protected] published 2026-11-10. The README says it does not connect to a live server, does not execute tools, and holds no conformance certificate.
- Offline
- 18 skills
- Node 20+
- npm latest
- [email protected]
- GitHub version
- 0.2.1
- Licence
- Apache-2.0
- Language
- HTML
- Weekly downloads
- 530
- Overall downloads
- 530
- GitHub stars
- 4
Live npm weekly downloads, npm overall downloads, and the latest npm version, plus GitHub stars and the package.json version, fetched 2026-11-17 19:07 UTC. Cached 10 minutes. Weekly downloads cover 2026-11-10 to 2026-11-16. Overall downloads cover 2026-10-01 to 2026-11-17. npm does not publish an earlier total. Stars and downloads show public use. They are not a quality score.
Use it when
- 1
You have an MCP tool snapshot or an A2A Agent Card saved as a file.
- 2
You want a heuristic scan for tool-poisoning text, hidden Unicode, and leaked secrets.
- 3
You want a CI gate where the worst result wins.
What it does, and what it does not
Read this before the guide. A limit here is a limit of the tool.
Does
- validate checks structure of MCP snapshots and A2A Agent Cards.
- audit is a heuristic scan for tool poisoning, secrets, high-impact tool names, and unsafe schemas.
- diff reports breaking changes between two MCP snapshots. replay reads a recorded JSON-RPC trace.
- gate combines reports and the worst result wins. Exit 2 is NO_GO or DENY. Exit 3 is REVIEW or INCONCLUSIVE with --strict.
- 18 skills copied into project folders. The README says IDE activation is not verified.
- The README records 121 passing tests. On 32 labelled descriptions, recall was 12 of 16 malicious (75%) with 1 false alarm in 16 benign. It says the 4 misses are deliberate evasions and the corpus is small and synthetic.
Does not
- It does not contact a live MCP server or A2A agent and does not execute tools.
- It is not an official conformance suite and is not affiliated with the MCP or A2A projects.
- PASS means nothing blocking was found by these checks. The README says that is never secure or compliant.
- Detection is heuristic. Paraphrase, other languages, and encodings evade it.
- Live MCP slices, a sandboxed runner, public-key signatures, SARIF, and a dashboard are planned, not claimed.
- 0 runtime dependencies. Node.js 20 or newer.
Claims above are from the project pages and are attributed as stated there. They are not independently verified here.
How a run moves
From the ProtocolSutra AI README, npm 0.2.0, published 2026-11-10. It evaluates captured files. It does not connect to a live server.
-
1. Input
A snapshot or trace
An MCP tool snapshot, an A2A Agent Card, or a recorded JSON-RPC trace. Not a live connection.
-
2. Check
Structure, then heuristics
validate is structural. audit looks for hostile tool text, hidden Unicode, secrets, and unsafe schemas.
-
3. Diff
Two snapshots
Breaking changes between MCP tool contracts. replay reads a recording. It does not execute the tools.
-
4. Gate
Worst result wins
Exit 2 is NO_GO or DENY. PASS means these checks found nothing blocking. It is not a conformance certificate.
How to read the result
Wording is from the project page or README and is attributed as stated there. It is not independently verified here.
Commands and notes
-
npx [email protected] fixture create --kind mcp --out snapshot.jsonsynthetic fixture, no credentials. - protocolsutra validate --input snapshot.json
- protocolsutra audit --input snapshot.json --format all --out audit-report
-
protocolsutra diff, replay, gate, and capsule export/verifylisted in the README command table. -
protocolsutra install-skillscopies 18 skills. The README says IDE activation is not verified.
Guide
How to use the result. GitHub is the source repository. npm is the install.
What the lab checks
From the ProtocolSutra AI README, npm 0.2.0, published 2026-11-10. Captured files only. Node.js 20 or newer.
- validate checks structure of MCP snapshots and A2A Agent Cards. audit is a heuristic scan for tool poisoning, secrets, high-impact tool names, and unsafe schemas.
- diff reports breaking changes between two MCP snapshots. replay reads a recorded JSON-RPC trace and does not execute the tools.
- gate combines reports and the worst result wins. Exit 0 is ok. Exit 1 is usage or runtime. Exit 2 is NO_GO or DENY. Exit 3 is REVIEW or INCONCLUSIVE with --strict.
- The poisoned fixture in the README is flagged for hidden instructions, hidden Unicode, and secret-access language. Reports: audit-report.json, .md, and .html.
What PASS is not
- It does not contact a live MCP server or A2A agent and does not execute tools. It is not affiliated with the MCP or A2A projects and holds no conformance certificate.
- PASS means nothing blocking was found by these checks. The README says that is never secure or compliant. Detection is heuristic and can be evaded.
- On 32 labelled descriptions, recall was 12 of 16 malicious (75%) with 1 false alarm in 16 benign. The README says the 4 misses are deliberate evasions and the corpus is small and synthetic.
- 18 skills. IDE activation is not verified. Live MCP slices, a sandboxed runner, public-key signatures, SARIF, and a dashboard are planned, not claimed. 0 runtime dependencies.
Documentation
GitHub Pages for this repository returned HTTP 404 on 2026-11-17. These cards link Markdown in the repository docs folder.
Questions
Does ProtocolSutra talk to my MCP server?
No. The README says it evaluates captured files. It does not connect to live servers and does not execute tools.
Is PASS a conformance certificate?
No. The README says PASS means nothing blocking was found by these checks, never secure or compliant, and the project holds no MCP or A2A certification.
Other projects
- VigilSutra AI 256 last week · 2,190 overall · 4 stars
- RachnaSutra AI 208 last week · 925 overall · 4 stars
- ArchSutra AI 176 last week · 883 overall · 4 stars
- PramanaSutra AI 195 last week · 750 overall · 4 stars
- EvalSutra AI 207 last week · 563 overall · 4 stars
- DocSutra AI 511 last week · 511 overall · 4 stars
- IncidentSutra AI 490 last week · 490 overall · 4 stars
- ContextSutra AI 529 last week · 529 overall · 4 stars
- MigraSutra AI 519 last week · 519 overall · 4 stars
- WattSutra AI 440 last week · 440 overall · 4 stars
- EdgeSutra AI 337 last week · 337 overall · 4 stars
- ContribSutra AI 274 last week · 274 overall · 4 stars
- LocalSutra AI 295 last week · 295 overall · 4 stars
- LineageSutra AI 0 last week · 0 overall · 0 stars
- DeploySutra AI 0 last week · 0 overall · 0 stars