What HolmesGPT/holmesgpt shipped

Holmes 0.46.0 is out: pick which MCP server tools load, send alerts to a webhook, and run AWS MCP from one hosted toolset across all accounts. Also in: image CVE fixes, a Prometheus null-timeout fix, and tool call status in logs.

Holmes 0.46.0 is out. You can now load only selected tools from an MCP server with allowed_tools, send alerts to a webhook destination, and serve all AWS accounts from one hosted AWS MCP toolset instead of one server per account. This week also fixed open CVEs in the holmes image by raising the aiohttp and certifi floors, fixed Prometheus instant and range queries failing on a null timeout, and added tool result status to the finished log line so success, error and no-data rates can be measured. A mounted kubeconfig is now refused unless bash is off in the toolsets Holmes receives, and k8sRBAC and mcpAddons changes now roll the Holmes pod.

Rate-limit and 503 errors are now retried with backoff instead of failing a run. Remote tool calls keep returned images. Holmes status reports each model's limits and their source. The weekly benchmark ran all 198 runs with no setup failures.

This week's changes to HolmesGPT: Provider rate-limit, overload and 503 errors are retried with backoff instead of failing the whole run and losing tool work already done. Remote tool calls keep the images a tool returned in the payload, so a rendered panel from another cluster arrives as an image rather than text. The toolset subtype is forwarded to multi-instance children, so prometheus/metrics honors subtype settings. Holmes status now reports each model's context and output limits and where they came from, instead of silently falling back to 200K context and 64K output. Each remote tool publishes a contract hash and each toolset row reports its Holmes version, so agents on older versions with identical arguments still join multi-agent investigations during a gradual upgrade. The weekly

This week: AWS MCP now uses the hosted AWS MCP Server, docs build with mkdocs-material 9.7.7 (search XSS fix), a webhook alert destination for health checks, sanitized causes kept on failed runs, and less duplicate data and repeated logs.

0.45.0 is out. New multi-tenant Prometheus toolset for Mimir, allowed_tools filtering for MCP servers, namespace/label/field selectors on the Kubernetes query tools, chunked history compaction, expired-JWT retries and patched CVEs in the image.

HolmesGPT 0.45.0 is out. This release adds a multi-tenant Prometheus toolset for backends such as Grafana Mimir, where the tenant is chosen by a Kubernetes label. MCP server entries now take allowed_tools patterns so only the tools you list are loaded. The Kubernetes query tools accept namespace, label and field selectors, which keeps investigations on large clusters from listing every namespace. Oversized conversation history is now summarized in chunks, PostgREST calls with an expired JWT are retried with new credentials, and the server reuses its cached tool executor in /api/info and the OAuth callback. The holmes image also ships patched critical and high CVEs.

Holmes 0.43.0 is out. Fast mode is now the default, the auto executor runs one conversation at a time, and conversations run on named executors for manual chats and auto triage. Per-user Dex auth for the Kubernetes MCP addon is documented. Details in the repo.

Holmes 0.43.0 is out this week. Fast mode is now the default and TodoWrite planning is opt-in. The auto executor defaults to one conversation at a time, so background triage, bulk investigations and triggered workflows no longer run concurrently, and conversations now run on named executors that separate live user chats from alert-triage work. The Kubernetes MCP addon gained per-user Dex auth with private CA support, with setup documented on the MCP page. On the reliability side, the model cost map refreshes on a lookup miss so recently released models are priced while Holmes is running, Robusta models whose upstream name is not provider/id are now priced, and free-form object parameters are kept out of strict tool calls.

This week: fixed a security issue in /eval and /list endpoints, removed the broken kubectl-run toolset, fixed SQL Server TLS validation, and standardized deployment documentation across 40+ integration pages.

This week's updates to HolmesGPT focused on security and documentation reliability. We fixed a critical issue where /eval and /list endpoints weren't properly gated on repository permissions, removed the kubectl-run toolset after consistent failures, resolved TLS certificate validation problems with pyOpenSSL 26.2, and standardized deployment instructions across 40+ integration pages to ensure readers get consistent, working steps every time.

Bash approval rules tightened for complex syntax. Six CVEs patched. Account opt-out settings now honored. kubernetes-remediation MCP 1.4.0. Bash docs clarified.

This week's HolmesGPT updates focus on security and reliability. Bash command execution now requires approval for complex syntax patterns. Six Critical/High CVEs in dependencies were resolved. The agent respects account-level opt-out settings for hosted models. kubernetes-remediation MCP upgraded to 1.4.0. Bash toolset documentation improved with real prefix examples.

0.42.0 ships security fixes for command injection and privilege escalation, hardened file-read policies, GPU diagnostics, and corrected failure analysis to constrain claims to evidence.

0.42.0 is out. This release hardens security with fixes for command injection and privilege escalation vulnerabilities, tightens file-read policy defaults, adds GPU node diagnostics, supports namespace-scoped deployments, and improves failure narratives by constraining claims to observed evidence only.