A collection of MCP server plugins for science, firebase, chrome-devtools, and data analytics capabilities including QuickGO, UniProt, BigQuery, and other specialized tools
This MCP server exhibits pervasive definition quality issues across 26 tools. While naming follows verb_noun conventions reasonably well, descriptions are often generic or missing critical context, input schemas lack required type information and parameter descriptions, and output schemas are not documented. The tool definitions appear inferred from Python scripts rather than explicitly registered with structured metadata. Many tools require parameters but provide minimal guidance on valid values, formats, or constraints. Error handling is absent, no recovery guidance or actionable error messages. The architecture mixes domain concerns (QuickGO, UniProt, Firebase, AlphaGenome, Chrome DevTools, BigQuery) without clear composition or chaining patterns.
Tools (26)
addCrashlyticsRunScriptBuildPhasewrite40/100
Add Crashlytics run script build phase to Xcode project
annotation_searchread onlysource verified58/100
Searches for annotations using the QuickGO API
bigquery_data_source_inspectread onlyauth50/100
Retrieve Data Source parameters for BigQuery Data Transfer Service or list available data sources
chat_expressread onlyauth50/100
Express.js chat endpoint that streams responses from the Gemini DataChat API
chat_fastapiread onlyauth50/100
FastAPI chat endpoint that streams responses from the Gemini DataChat API
compare_snapshotsread only50/100
Compare two Chrome DevTools heap snapshots and identify growing objects
Tool definitions are inferred from Python scripts rather than explicitly registered with structured MCP metadata. No explicit tool registration with full JSON Schema is visible in the source code.
Explicitly register all tools with the MCP server using the tools/list response format, including full JSON Schema input schemas with $schema, title, type, properties, required array, and descriptions for each parameter.
Add complete output schemas to every tool definition. Specify the response structure, field types, and constraints. Example for go_search: 'Returns an object {results: [{id: string, name: string, description: string}], total: number, page: number}'.
Expand tool descriptions to 50 - 200 characters. Include WHAT the tool does, WHEN to call it (vs. similar tools), and any prerequisites. Example: 'Search Gene Ontology (GO) terms by text query. Use this to discover GO term IDs for annotation queries. Returns matching terms with IDs and definitions. For exact GO ID details, use go_terms instead.'
Add descriptions to every input parameter. For parameters with constrained values (e.g., 'relation' in go_terms), use enum instead of free-form strings. For 'ids' parameters accepting comma-separated values, clarify format: 'Comma-separated GO IDs (e.g., GO:0006915,GO:0008150)'. For 'limit' parameters, declare min/max: 'Max results per page (1 - 100, default 25)'.
Remove 'output' as a required parameter from tools. Instead, return results directly in the tool response. The MCP server/client handles persistence, not the agent.
Document parameter relationships and dependencies. For annotation_search, clarify that 'goUsage' only applies when 'goId' is provided. Use conditional parameter language: 'Required if goId is specified'.
Fetches AlphaGenome metadata and writes the ontology mapping JSON
get_countread onlysource verified57/100
Retrieve the total number of hits for a query
get_entryread onlysource verified62/100
Retrieve a single UniProt entry
get_project_idread onlyauthsource verified47/100
Retrieves the active Google Cloud project ID
get_regionread onlyauthsource verified45/100
Retrieves the default compute region from gcloud config
get_tokenread onlyauthsource verified38/100
Retrieves the access token using the gcloud CLI
go_searchread onlysource verified60/100
Search GO terms by text
go_slimread onlysource verified45/100
Calculate GO Slims
go_termsread onlysource verified53/100
Get GO term details
hasCrashlyticsRunScriptBuildPhaseread only37/100
Check if Crashlytics run script build phase exists in Xcode project
isUserScriptSandboxingEnabledread only33/100
Check if user script sandboxing is enabled in Xcode project
parseSnapshotread only50/100
Parse Chrome DevTools heap snapshot file and extract node statistics
run_id_mappingread onlysource verified57/100
Execute the ID mapping workflow to map protein IDs between different databases
search_proteinsread onlysource verified65/100
Search proteins in a UniProt dataset with automatic pagination
setDwarfWithDsymDebugInformationFormatwrite37/100
Set DWARF with dSYM debug information format in Xcode build configurations
sparql_queryread onlysource verified45/100
Execute a SPARQL query
stream_resultsread onlysource verified62/100
Stream all results for a bulk query using the /stream endpoint
Input schemas lack proper JSON Schema type declarations and descriptions for many parameters. Parameters are visible but types are inferred, not explicitly declared. For example, 'slimsFromIds' and 'slimsToIds' in go_slim lack min/max constraints or enum values.
Output schemas are not documented. Tools like go_search, annotation_search, and search_proteins do not declare what fields they return or their types. LLMs cannot plan downstream operations without knowing output structure.
Tool descriptions are generic or incomplete. Many descriptions lack context on WHEN to use the tool vs similar ones, or lack prerequisites. Examples: 'Search GO terms by text' (go_search) does not explain the difference between go_search and geneproduct_search. 'Check if user script sandboxing is enabled' (isUserScriptSandboxingEnabled) lacks explanation of what to do if enabled/disabled.
No error handling or recovery guidance. Tools provide no error messages or guidance for LLM recovery. If a search returns no results, the API error is not documented. If an ID mapping fails, no actionable message tells the agent what to try next.
Parameter naming lacks type suffixes. Parameters like 'ids' in go_terms accept comma-separated strings but do not specify format (e.g., 'go_ids' or 'comma_separated_go_ids'). Similarly, 'query' in search_proteins and sparql_query are ambiguous, they could accept natural language, SQL, SPARQL, or regex without clear guidance.
Parameter limits are undocumented or inconsistent. Many search tools mention 'max 100' results inline but do not expose it as a constraint in schema (min/max). The 'limit' parameter in go_search defaults to 25 but nowhere is it documented whether 0 is valid or if pagination is required for large result sets.
Many tools expose 'output' as a required parameter for writing results to files. This is inappropriate for an MCP tool interface, output file handling should be server-side or handled by the client, not exposed as a parameter. Requiring agents to specify file paths couples the agent to the filesystem and leaks implementation details.
Firebase Xcode tools (isUserScriptSandboxingEnabled, hasCrashlyticsRunScriptBuildPhase, etc.) accept 'project' as an object parameter but do not document what fields are required, their types, or valid values. Passing a PBXProj object to an MCP tool is non-standard and unclear.
Tool composition is unclear. The 26 tools span 6+ domains (QuickGO, UniProt, Firebase, AlphaGenome, Chrome DevTools, BigQuery, DataChat) with no tool chaining documented. There is no guidance on which tools compose (e.g., should go_search be followed by go_terms? should search_proteins be followed by run_id_mapping?). Output from one tool rarely includes IDs needed by the next.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools like addCrashlyticsRunScriptBuildPhase and setDwarfWithDsymDebugInformationFormat are write operations but lack destructiveHint to signal irreversibility to the LLM. Tools like get_project_id and get_token are read-only but lack readOnlyHint.
Add error handling and recovery guidance to tool docstrings. Example: 'If no results found, verify spelling of GO ID or try go_search with a text query instead.'
Add tool annotations. Mark write tools with 'destructiveHint: true' and read-only tools with 'readOnlyHint: true'. Mark idempotent operations (get_entry, get_project_id) with 'idempotentHint: true'.
For Firebase Xcode tools, replace the 'project: object' parameter with explicit string/boolean parameters. Clarify what a 'PBXProj object' is in the context of MCP, likely a file path or serialized JSON structure. Document the required fields.
Document tool chains. Add a tool_composition field or description section explaining which tools typically sequence. Example: 'go_search finds IDs → go_terms retrieves details → annotation_search finds annotations using those IDs'.
Validate all numeric parameters early. For 'limit' parameters, enforce 1 - 100 and return a clear error: 'Invalid limit: got 500, must be 1 - 100'. For 'page', ensure page > 0.
For SPARQL and query-based tools, document acceptable query syntax or provide examples. 'sparql_query expects a valid SPARQL SELECT query. Example: SELECT ?protein WHERE { ?protein a <http://example.org/Protein> }'. Return clear errors if the query is invalid rather than dumping API stack traces.
Unify pagination patterns. All search tools should support 'page'/'limit' OR 'offset'/'limit', not both, not mixed. Document whether pagination is 0-indexed or 1-indexed.
For tools that call external APIs (QuickGO, UniProt, BigQuery), add timeout guidance and retry hints. 'If the request times out, retry once before notifying the user.'
Add batch variants for tools called in loops. Instead of agent calling add_label 10 times, offer add_labels accepting an array. This saves context and latency.
Clarify the difference between similar tools. Examples: 'go_search' (text search in GO) vs 'geneproduct_search' (search for proteins/genes, not ontology terms)', 'search_proteins' vs 'get_entry' (bulk search vs. single lookup).
Document what each tool returns on success and partial failure. If run_id_mapping succeeds for 8 of 10 IDs, return {successful: [...], failed: [...], errors: {...}} rather than a flat list or an all-or-nothing error.