MCP directory preflight
Compare labelled tools/list exports locally for raw and client-visible name collisions, missing titles and boolean hints. Includes pinned client profiles.
Runs entirely in your browser
Free, browser-local metadata checks. No URL fetching, sign-in, tool calls or automatic fixes. Use sanitized JSON without private account data.
Local report
Presence and type checks only. A declared hint does not prove semantic safety or directory acceptance.
Metadata issues and scope
Supplied tool names and explicit hints
256 KiB of UTF-8 JSON and labels, 200 tools total, 64 nesting levels. Pasted or picked exports only. No server, authentication or tool calls; no nested JSON Schema validation.
Compare tools/list exports
- Give each export its server name as a source label, then paste tools/list JSON or choose a local file. Add exports to compare servers or supplied pages.
- Run the local check. Inspect collision groups and client-visible names; each row identifies its source, export and original JSON Pointer.
- Review actual tool behavior and directory requirements yourself. Download the report if useful; clear the page when finished.
What this checks
Supply one or more {"tools":[...]} exports or JSON-RPC 2.0 responses with result.tools. Give each its actual configured server name as a source label; reuse that label for pages from the same server. Labels affect qualified names and are never inferred from filenames. Arrays, requests, error responses and duplicate JSON members are refused. A UTF-8 byte-order mark is accepted. A supplied nextCursor triggers a partial-list warning; the page never fetches another page.
Client-visible name checks
Raw duplicates are grouped across all supplied exports. Each client view shows its transformed names and collision groups, with source labels and JSON Pointers. These are naming snapshots, not universal client compatibility tests: other versions, configured overrides and locally defined non-MCP tools are not checked. Late-loaded or OAuth-gated tools and pagination beyond the pasted pages remain unchecked.
OpenAI Agents JS, checked 2026-11-14: utils/tools.ts at fdaf0a6 replaces non-alphanumeric characters with underscores in default unprefixed mode; mcp.ts rejects resulting duplicates after PR #1935. For example, search-a and search_a collide. Its opt-in server-prefix mode is not modeled.
Gemini CLI, checked 2026-11-14: mcp-tool.ts in v0.68.0 (5a5c718) builds the default mcp_source_tool name, replaces unsupported characters, then shortens names over 63 characters to the first 30, ... and last 30. Configured name overrides are excluded.
Separate provider rule, checked 2026-11-14: OpenAI Chat Completions function.name allows ASCII letters, digits, underscores and hyphens, up to 64 characters. This page flags names outside that rule without renaming them; Gemini allows other characters. No automatic renaming or tool calls occur.
Each result gives JSON Pointer paths, such as /result/tools/0/annotations/readOnlyHint. Checks cover names, duplicate names, nonblank descriptions, object-shaped input and output schemas, title fields and the types of the four known boolean hints. Nested JSON Schema constraints, descriptions, permissions and actual tool behavior need separate review.
Two dated directory profiles
OpenAI profile checked : its submission guidelines require explicit readOnlyHint, openWorldHint and destructiveHint booleans on every tool. No values are inferred.
Anthropic profile checked : its review criteria require a title and the applicable hint: readOnlyHint: true for reads and destructiveHint: true for tools that modify or delete data. Its directory policy calls for title annotations, so this profile checks annotations.title. You must decide which hint applies by inspecting behavior.
The MCP specification treats annotations as untrusted hints. Its display title can come from title or annotations.title; protocol defaults do not meet a directory's explicit-field requirement. Directory policies can change after the dates above. Follow the linked requirements before submitting.
These tools are open source (MIT).
For a real remote connectivity, authentication and tool-list probe, use the paid agentscan remote MCP check: POST /v1/mcp/check.
Questions
Is the JSON uploaded or saved?
No. Input processing stays in this tab, with no processing requests or saved inputs. Choosing a file reads only that file. Clear removes the input and report; closing the tab discards page memory. A report you download remains on your device. Normal page assets load first; page-view analytics never receives your JSON.
Does a clean report mean directory approval?
No. Hints are untrusted declarations. This page does not inspect code, invoke tools, test OAuth or decide whether an annotation matches behavior. Each directory applies its own current policies and review.
Why does the Anthropic profile ask for annotations.title?
Its directory policy asks for title annotations. MCP also allows a top-level tool title, with display precedence over annotations.title. This profile checks the annotation separately and does not treat the tool name as a title.
Are omitted hints filled in or JSON files changed?
No. The check preserves explicit true and false values and reports missing or mistyped values. It never assigns MCP defaults, rewrites annotations or edits a file. Duplicate JSON members are rejected instead of keeping the last value.