Claude Code

The Synthient plugin for Claude Code: the MCP server for data, and six skills for the judgment the tools cannot carry.

Install

/plugin marketplace add synthient/plugins
/plugin install synthient@synthient

Claude Code prompts for an API key when the plugin is enabled. Leaving that blank is fine. The bundled MCP server falls back to the CLI's own credential chain: SYNTHIENT_API_KEY, a .env in the working directory, then the OS keychain populated by synthient auth. Get a key from the dashboard at synthient.com.

synthient auth

Confirm everything resolved with /synthient:doctor, which walks the CLI install, the active credential source, the granted scopes, remaining credits, and MCP connectivity, stopping at the first failure.

Skills

SkillPurpose
/synthient:docsAPI reference: auth, response shape, field vocabularies, credits, rate limits, errors
/synthient:integrateWire Synthient into the current codebase, with fail-open handling and the benign-automation guard
/synthient:migrateMove off Spur, IPQualityScore, or another vendor: audit, map, report, rewrite, verify
/synthient:triageInvestigate an address, a list, or a domain and return a reasoned verdict
/synthient:feedsParquet snapshots, filtered stream capture, CSV export, local analysis
/synthient:doctorDiagnose credentials, scopes, credits, and MCP connectivity

Claude loads these on its own when the context calls for it: a Synthient field name in a diff, a 403 from api.synthient.com, an ask to check an address. The slash commands are for when you want one specifically.

Migrating from another vendor

/synthient:migrate spur

Also takes ipqs, or any other vendor name. A read-only subagent audits every call site first, and the report lands before anything is rewritten.

Vendors without a published guide go through a derived mapping, marked as inferred.

See Migrating from Spur and Migrating from IPQualityScore for the mappings it works from.

What it adds over the tools

The skills supply the judgment: that nine of the intelligence.categories values are benign automation, where blocking takes out Googlebot and Stripe webhooks; that network.type is a network classification and not a proxy verdict; that risk_score is a summary rather than a verdict; that timestamp units differ by surface, with the anonymizer and Helios HTTP, HTTPS, and DNS streams in milliseconds and lookups, exports, and the other streams in seconds; that an empty providers[] is not a clean bill of health; that rate limits are per team, so a second key does not double throughput.

They also reach what the tools cannot: Parquet downloads with checksum verification, filtered and long-running stream capture, CSV export, and status / scopes diagnostics. Reference detail is fetched live from these docs.