Changelog

0.7.0

Breaking

• The HTTP transport now requires inbound authentication. Docker and other MCP_TRANSPORT=http deployments refuse to start until MCP_AUTH_JWKS_URI, MCP_AUTH_ISSUER and MCP_AUTH_AUDIENCE are set, rather than serving vulnerability data to anyone who can reach the port. docker-compose.yml passes the three variables through. Stdio is unaffected.

Added

• Inbound authentication for remote mode. The HTTP transport validates OIDC bearer tokens against any standards-compliant identity provider, configured from environment (MCP_AUTH_JWKS_URI, MCP_AUTH_ISSUER, MCP_AUTH_AUDIENCE) and read at startup with no image rebuild. Several issuers can be accepted at once. Stdio remains unauthenticated by design. See docs/authentication.md.
• Read/write scope separation. The mutating tools — start_rapid7_export, download_rapid7_export, load_rapid7_parquet and purge_rapid7_data — require a write scope (MCP_AUTH_WRITE_SCOPE, default rapid7.write) on the HTTP transport; read tools stay available to any authenticated caller. Gating is inert on stdio.
• rapid7-refresh foreground entrypoint. A headless console script that creates, polls, downloads and loads the requested exports synchronously in a single process — no background threads that could die mid-write — for scheduled and hosted refresh. Exits non-zero and names the failed windows if any window fails. In hosted mode it publishes the finished database to Blob Storage as a versioned artifact.
• Data-age annotation on hosted query results. When serving a Blob artifact, successful query_rapid7 responses carry a short data-age note sourced from a load-metadata table inside the data database. It is fail-soft and never breaks a query. Local output is unchanged.
• Optional query time limit. DUCKDB_QUERY_TIMEOUT_SECONDS cancels a query that runs longer than the limit with an actionable message. Unset by default, so local queries run to completion as before; the Azure template sets 90 seconds.
• Private hosting guide for Microsoft Copilot Studio. See docs/copilot-studio-hosting.md for an internal-ingress deployment with Blob artifact storage, a scheduled refresh job, and Key Vault–backed secrets.

0.6.3

Fixed

• Export status and download work again. v0.6.2 declared the export id in the status query as ID!. The Rapid7 export schema has no ID type, so every status and download call failed with Unknown type 'ID'. The query now uses String!. Export creation was not affected.

Changed

• Bumped pyjwt 2.13.0 → 2.15.0 and urllib3 2.7.0 → 2.8.0.

0.6.1

Changed

• Adopted the Agent Plugins packaging format. The repository now ships a root plugin.json + mcp.json and a skills/rapid7-bulk-export/ directory. Kiro and other conformant clients install the MCP server and skill together as one power. The manifest.json (MCPB) bundle for Claude Desktop connector install is unchanged.
• Removed the legacy power-rapid7-bulk-export/ (POWER.md) layout, superseded by the Agent Plugins format for the same client. Existing installs of the old-format power keep working until they are re-pulled; re-install from the repository URL to move to the new format.

Added

• PLUGIN_DATA support for the database location. When DATA_DIR is not set, the server now uses the plugin host's PLUGIN_DATA directory (per-install, writable, survives updates) before falling back to ~/.rapid7_mcp. An explicit DATA_DIR still takes precedence, so existing installs are unaffected. This also benefits MCPB installs.

Fixed

• Reconciled the skill's tool references to the real prefixed tool names (query_rapid7, load_rapid7_parquet) so documentation matches the server.

0.6.0

Fixed

• Multi-month remediation exports now load every window. Previously, requesting remediation data for a range longer than one month could silently load only a single 31-day window. The Rapid7 platform allows only one remediation export in flight at a time and splits a longer range into multiple exports; the tool created them without sequencing, so all windows collapsed onto one export ID and only one month's data was downloaded — and it could be any month in the requested range.

  Impact: if you ran a remediation export covering more than 31 days on an earlier version, it may have loaded only part of your range. Re-run that export on 0.6.0 to get the complete data.

Changed

• Requesting a remediation range now starts a single background job that creates, downloads, and loads each ≤31-day window in order and appends them into one vulnerability_remediation table. start_rapid7_export(export_type="remediation", …) returns a job ID; poll it with check_rapid7_export_status. Partial failures are explicit — the job reports exactly which windows loaded and which are missing so the missing range can be re-run.