Valyd
Discover what changed across the MCP ecosystem.
Valyd publishes what MCP servers expose and how it changes from release to release: every tool, its schema and its declared authority, with evidence-backed comparisons between versions. Explore it in your browser. There is nothing to install and no account to create.
- Public and free, no account
- No CLI install needed
- Agent Interface Intelligence by Modelara
Filesystem MCP · 7 releases analyzed
- 2.1.0 → 2.2.0FAIL2 added · 3 schema · 2 writePending human review
- 2.0.1 → 2.1.0PASS0 added · 0 schema · 0 writeAutomated static analysis
- 2.0.0 → 2.0.1PASS0 added · 0 schema · 0 writeAutomated static analysis
- 2026.8.31 → 2.0.0REVIEW0 added · 14 schema · 0 writePending human review
- 2026.7.10 → 2026.8.31PASS0 added · 0 schema · 0 writeAutomated static analysis
- 2026.7.4 → 2026.7.10REVIEW0 added · 1 schema · 0 writePending human review
01 The problem
A name and a version number don’t tell you what an agent can do.
Choosing or upgrading an MCP server usually starts with its package name, description and version. None of them says which tools an agent will be given, what arguments those tools accept, or what changed since the version you run. Here is what the Filesystem MCP server’s own history shows.
| What you can see | What you need to know | What Valyd published |
|---|---|---|
Version number2.1.0 → 2.2.0, a minor release | Whether new capabilities appear | Two new tools that write files: append_file and copy_file |
Version scheme2026.8.31 → 2.0.0 | How parameters change | 14 schema changes, published as REVIEW |
Package name and description@modelcontextprotocol/server-filesystem | What tools are exposed | 16 tools in 2.2.0, each with its input schema and effect |
Registry version listEvery published version | Which versions were actually observed | 7 releases analyzed, compared as 6 consecutive transitions |
| A claim that “nothing breaking changed” | What evidence supports a comparison | Analyzer version, package integrity hash and a fingerprint for every tool |
02 Real interface evidence
Filesystem MCP 2.1.0 → 2.2.0, as published.
A minor release that adds two tools able to write files and changes three input schemas. Everything below is transcribed from the published comparison (opens in a new tab) and the server’s evidence page (opens in a new tab).
Filesystem MCP
2.1.0→2.2.0
- Tools
- 14 → 16
- Tools added
- 2
- Input-schema changes
- 3
- Description changes
- 3
- New write authority
- 2
append_filenew tool · authority — → WRITEFAILcopy_filenew tool · authority — → WRITEFAILdirectory_treeinput property added: depthREVIEWread_filehead, tail: number → integerREVIEW
Two added tools that write
+ append_fileFAIL- Effect
- write (read_only: false · destructive: false)
- Inputs
- path, content (both required)
- Fingerprint
sha256:ecc921d116b1b42f8e30e529591961dfca69c56210e880887669b2ca63495cf0- Interface hash
sha256:a5bbc742948507b1fa54b2455343ad7395b6f92cad36adfd83f953840b7576c0
+ copy_fileFAIL- Effect
- write (read_only: false · destructive: false)
- Inputs
- source, destination (both required)
- Fingerprint
sha256:8344300e46804c5780de927f2cc7164ffacb9495c857be6ea1ce7400b8a85cc4- Interface hash
sha256:bc6b70190798aff499fc9e6be26469987f9c7e96c7b4550f85be1dd7ac1f1d6d
Input schema changes
~ directory_treeinput property added: depthREVIEW~ read_filehead, tail: type "number" → "integer", +minimum=0, +maximum=9007199254740991REVIEW~ read_text_filehead, tail: type "number" → "integer", +minimum=0, +maximum=9007199254740991REVIEW
Provenance for 2.2.0
- Package
@modelcontextprotocol/server-filesystem@2.2.0- Registry
npm- Published
Nov 4, 2026- Integrity
sha512-mNLj/qhV+P4kj1qP8VzgQ2fjSmWcKDQ3Q+KpZEA+WahCDWHw5zJBRQUNSPZ6Ai2/ZGGUjzNzsFpz958a1+4T9A==- Analyzer
modelara v0.1.8 · policy not recorded- Repository
github.com/modelcontextprotocol/servers
The comparison is FAIL because the policy requires a person to approve new write authority. It is not a security score, and it does not mean the release is unsafe. It is published as Pending human review: the analysis is automated and no reviewer decision has been recorded for it yet. Ten informational metadata changes (declared idempotent: false → true) are on the comparison page and not repeated here.
03 What Valyd gives you
More than a directory: observations, comparisons and evidence.
A directory tells you a server exists. Valyd records what each release exposes, compares it with the release before, classifies what changed and keeps the evidence for every result.
MCP server discovery
Browse tracked servers by category and by distribution: npm, PyPI, Go, OCI / Docker, MCPB. Each server page names its package and links to its source.
Public catalog
The latest comparison for every tracked server in one table. Filter by the kind of change: write authority, destructive authority, unknown effect, schema changes, removed tools.
Version observations
A release history per server, and a capability surface showing the first version each tool was observed in Valyd’s coverage and the version where it last changed.
Interface comparisons
Each pair of consecutive stable releases is compared statically: tools added and removed, input and output schema changes, description and metadata changes.
Tool and capability changes
Every capability is classified Read, Write, Destructive or Unknown, labelled declared or inferred. Unknown is never treated as safe. New authority is counted separately.
Evidence and provenance
Observed schemas with their own SHA-256 hashes, per-tool fingerprints, package integrity, publication time, repository and the analyzer version behind each result.
Change classifications
Every comparison is PASS, REVIEW, FAIL or INVALID under Modelara’s published policy, and each finding is grouped as a policy failure, requires review, or informational.
Watchlist and change feed
Watch the servers your agents use, saved in your browser with no account, and see their changes in one place. Every change is also in a public JSON feed.
Server and comparison counts change with each publication and are shown live on valyd.dev/mcp (opens in a new tab). The change feed is at /feeds/mcp-changes.v1.json (opens in a new tab).
04 How Valyd works
Nothing is published until it is reviewed.
Valyd’s data is produced by Modelara’s analysis and published through a controlled, reviewed process. The full rules are in the Valyd methodology (opens in a new tab).
- 01
Discover server sources
Catalog entries are reviewed and configuration-driven. A server is tracked only when its package identity is known, versions can be pinned and earlier versions are available.
- 02
Observe and inspect releases
Stable releases are ordered by semantic version; prereleases are excluded. Each one is started over stdio and inspected with a pinned Modelara CLI. Tools are listed, never called.
- 03
Record interface evidence
Each snapshot keeps the observed tools and schemas, canonical hashes, package version, publication time, integrity value and repository, plus the analyzer version.
- 04
Compare interface versions
Consecutive releases are compared over the saved snapshots. No model, API key or task suite is involved.
modelara check --static --json - 05
Review the findings
New data is proposed for review, and a person reviews it before it reaches the site. New write, destructive or unknown-effect changes stay marked pending review.
- 06
Publish approved intelligence
Approved data is published as one versioned generation. Each deployment of valyd.dev names the generation it was built from, and earlier analyzer provenance is not silently rewritten.
Publication review approves what reaches the site. A comparison’s own review status is shown separately, so an automated result is never presented as human-verified. You can check which published generation the live site was built from in valyd-data-source.json (opens in a new tab).
05 Part of Modelara
Valyd tells you what changed. Modelara tells you whether it matters to you.
Valyd, Modelara CLI and Modelara Observatory share one focus, Agent Interface Intelligence, and serve different workflows. Valyd is where approved public results are published; the CLI and Observatory are where interfaces are inspected and evidence is produced.
Inspect locally
Modelara CLI
- Inspect MCP servers on your machine
- Compare the versions you choose
- Produce PASS, REVIEW and FAIL signals as text, JSON or SARIF
Observe and review
Modelara Observatory
- Track selected MCP sources
- Inspect eligible artifacts
- Preserve evidence and review decisions
- Support controlled publication
Publish and discover
Valyd
- Make approved public interface observations and comparisons discoverable
- Free to browse, with no install or account
“Whether it matters to you” means checking your own versions and your own approvals, for example with modelara check --static or a model-based modelara test of your tasks. No Modelara product decides that an upgrade is safe for every agent.
06 Who it is for
One public record, four ways to use it.
- MCP server developers
- See how each of your releases changed the interface consumers depend on: which tools appeared, which schemas tightened, what authority was added.
- AI agent developers
- Research an external MCP server before you integrate it, and read what a new release changed before you upgrade.
- Platform engineering teams
- Review published changes across the servers you allow, and decide which upgrades need your own testing first.
- Developers exploring MCP
- Discover what real servers expose: their tools, schemas, versions and where each package came from.
07 Trust, evidence and limitations
Precise about what it shows, and what it doesn’t.
Valyd’s value is that every result is specific and checkable. That precision is also its boundary.
- Observed interfaces
- Valyd reports what each server declared when it was inspected: tools, schemas and effect annotations. Effects are inferred only where name, description and schema agree.
- Tied to specific artifacts
- Every result names the exact package version it came from, with its integrity hash or digest and the analyzer version. It says nothing about versions it did not inspect.
- Static, not runtime
- Findings come from the declared interface. They do not show whether an agent workflow still works; that needs a behavioral test with your own tasks.
- A change is not a vulnerability
- A finding records that something changed. It is not a security score, and Valyd does not certify any server as secure or unsafe.
- FAIL is a policy outcome
- FAIL means new or stronger authority that the policy requires a person to approve. It is not evidence that a server is malicious or broken.
- Coverage is what was observed
- Only reviewed catalog entries and the versions actually inspected and published are covered, over stdio. HTTP transport is not included.
- Publication is controlled
- Results are reviewed before they are published, so Valyd can lag behind upstream releases. The newest upstream release may not be on Valyd yet.
- Private evidence stays private
- Observatory keeps evidence and review history that are not published. Only approved records appear on Valyd.
Explore Valyd.
Find the MCP servers your agents use and see what their last release changed, or start a watchlist (opens in a new tab) to follow them.