Menu
Use when the user wants to gather context across multiple read-only connectors (jira, github, db, notion, ...) from a single natural-language prompt. The hub plans + dispatches connector calls in parallel.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Use when the user asks about Confluence content or wants to create, update, or delete Confluence pages and comments — CQL search, page retrieval, space metadata, comments, attachments, or any write operation on pages. Supports both read and write surfaces; every action passes the policy gate before any network call is made: read actions succeed by default, write and delete actions escalate by default, privilege actions are hard-denied.
Use when the user asks about Jira data or wants to create, update, or delete Jira content — JQL search, issue details, project metadata, comments, attachments, status transitions, or any write operation on issues and comments. Supports both read and write surfaces; every action passes the policy gate before any network call is made: read actions succeed by default, write and delete actions escalate by default, privilege actions are hard-denied.
Use when the user asks about read-only GCP inventory — Cloud Run services, Cloud SQL instances, Pub/Sub topics, or Cloud Logging entries. Queries are scoped to a single GCP project. Never modifies GCP resources.
Use when the user asks about GitLab data or wants to act on GitLab — project info, code search, issue/MR/release/pipeline inspection, or any write that creates, updates, closes, or deletes an MR, issue, note, or release; merging MRs (admin); and managing CI pipeline runs (retry / cancel / trigger / play_job).
Use this skill whenever the user wants to add a new agent connector to their @narai connectors workspace — wrapping a SaaS API, REST endpoint, GraphQL endpoint, SDK, or CLI tool into the canonical @narai connector shape (CLI + library + Claude Code plugin). Trigger even if the user doesn't say "connector" explicitly: phrases like "I want to query Stripe from Claude", "add Slack to our agents", "wrap our internal orders API", "connect Salesforce", "I have a GraphQL endpoint we need exposed", "make a Linear agent" all warrant this skill. Walks the user through identity, auth, and action-surface questions, then scaffolds the package, plugin layer, and tests by stamping out canonical templates and generating per-connector code. Do NOT use for: modifying existing connectors (just edit the file), wrapping MCP servers (different abstraction), or querying databases (db-agent-connector already covers postgres/mysql/sqlite/mssql/mongodb/dynamodb/oracle).
Use when the user asks about read-only AWS inventory — Lambda functions, RDS databases, S3 buckets, or CloudWatch metrics in an AWS account. Queries are scoped to a single region (except list_buckets, which is account-wide). Never modifies AWS resources.
| name | connector-hub |
| description | Use when the user wants to gather context across multiple read-only connectors (jira, github, db, notion, ...) from a single natural-language prompt. The hub plans + dispatches connector calls in parallel. |
| context | fork |
Answer the user's question by invoking the connector-hub binary exposed by
this plugin. It delegates to @narai/connector-hub, which reads
~/.connectors/config.yaml + ./.connectors/config.yaml, asks the Claude
Agent SDK for a plan, and dispatches each step in parallel.
connector-hub --prompt "<user's request>" [--consumer <name>] [--environment <name>] [--extra-context "<text>"]
Return the JSON envelope verbatim. The consumer parses plan + results
themselves.
{
plan: [{ connector, action, params }, ...],
results: [
{ step, connector, action, params, envelope },
{ step, connector, action, params, error: { code, message } },
...
],
}
Errors (planner failures, malformed plan entries, dispatch failures) appear
inline in results[] keyed by error.code rather than throwing.
Read-only — every dispatched connector enforces its own read-only guardrails. The hub itself only orchestrates; it never writes.
See the package README for full API details.