Comparison — vendor pages read 2026-11-11
Observability Unified vs Grafana Cloud
Grafana Cloud hosts Grafana's observability stack, including Tempo for traces and Loki for logs, and the open-source components can also be self-hosted.
Observability Unified is a self-hosted platform that connects traces, logs, replay, profiles, and AI agent activity through shared IDs, and gives AI agents evidence with citations through MCP. Each row below says Yes, Partial, or No for both products and links to the page that supports it.
Where Observability Unified differs
- MCP write tools off unless enabled. Writes register only with OBS_MCP_ENABLE_WRITES=true
- MCP results link to source records. Evidence bundles with citations; dashboard deep links when OBS_DASHBOARD_URL is set
- Compacted evidence for AI agents. CCR bundles collapse repeated logs to exemplars within a token budget; refs expand them
- Records what an agent's tool call changed. Before, after, and diff values your app attaches to a tool call, stored as sent
- Compare releases and flag variants. Error-rate lift per version and flag variant, normalized for traffic share
- Fix verified by re-running the failing case. Same eval input fails on the baseline and passes on the candidate
- Remediation behind approval or a policy gate. Approval queue or single-use grant; your runner executes; a watchdog checks recovery
- Session replay. rrweb, browser DOM only
Where Grafana Cloud offers more
- Write your own queries over telemetry. TraceQL, LogQL and PromQL for traces, logs and metrics
- Vendor-hosted cloud service. Grafana Cloud
- Built-in single sign-on. OIDC in all editions; SAML in Enterprise and Cloud
Capability by capability
Investigating with AI agents
| Capability | Observability Unified | Grafana Cloud |
|---|---|---|
| First-party MCP server | Yes Local stdio server, @obsunified/mcp-server our docs | Yes Hosted Cloud server and a self-managed OSS server grafana.com |
| MCP write tools off unless enabled | Yes Writes register only with OBS_MCP_ENABLE_WRITES=true our docs | Partial Write access is a consent checkbox; the OSS server has --disable-write grafana.com |
| MCP results link to source records | Yes Evidence bundles with citations; dashboard deep links when OBS_DASHBOARD_URL is set our docs | Partial A generate_deeplink tool for dashboards and Explore grafana.com |
| Agent tool calls alongside app traces | Yes Agent Action Graph links tool calls to traces our docs | Yes Generations link to their traces in Tempo grafana.com |
| Compacted evidence for AI agents | Yes CCR bundles collapse repeated logs to exemplars within a token budget; refs expand them our docs | Partial MCP server returns Loki log patterns; log queries flag truncated results github.com |
| Agent skills for coding agents | Yes Instrument and investigate skills, released as .skill bundles our docs | Yes grafana/skills Agent Skills for Claude Code, Cursor and Codex github.com |
Observing your own AI agents
| Capability | Observability Unified | Grafana Cloud |
|---|---|---|
| Records what an agent's tool call changed | Yes Before, after, and diff values your app attaches to a tool call, stored as sent our docs | Partial Agent Observability shows tool arguments, results and status grafana.com |
| MCP calls recorded with trace context | Yes MCP spans join the action graph; an optional allow-listed envelope records the context our docs | Yes OpenLIT MCP Observability traces MCP clients and servers grafana.com |
| Production traces become eval cases | Yes Save as eval case from a trace or run; candidate runs compared step by step (HTTP API) our docs | Yes Promote conversations into test suites and compare runs grafana.com |
Change-aware investigation and remediation
| Capability | Observability Unified | Grafana Cloud |
|---|---|---|
| Compare releases and flag variants | Yes Error-rate lift per version and flag variant, normalized for traffic share our docs | Partial Group and filter by service.version grafana.com |
| Shared-dependency contention vs code regression | Yes Shared-degradation verdicts separate contention from the new version our docs | Yes Knowledge Graph root-cause analysis across dependencies grafana.com |
| Root cause states its level of proof | Yes Proof tier from unverified locus to counterfactual verified, kept apart from link certainty our docs | Yes Investigations show each hypothesis's status and evidence grafana.com |
| Fix verified by re-running the failing case | Yes Same eval input fails on the baseline and passes on the candidate our docs | Partial Sandboxes run checks on a proposed fix grafana.com |
| Remediation behind approval or a policy gate | Yes Approval queue or single-use grant; your runner executes; a watchdog checks recovery our docs | Partial Per-tool approval policies; PR publish approval grafana.com |
| Matches new incidents to past ones | Yes Prior findings matched by fingerprint appear in incident evidence bundles our docs | Yes Past incident lessons feed later investigations (private preview) grafana.com |
Signals
| Capability | Observability Unified | Grafana Cloud |
|---|---|---|
| Distributed tracing (OTLP) | Yes OTLP over HTTP; front gRPC with a Collector our docs | Yes Tempo with OTLP ingest grafana.com |
| Structured logs | Yes Trace-correlated our docs | Yes Loki grafana.com |
| AI / LLM observability | Yes AI calls, agent runs, evals our docs | Yes Agent Observability, GA July 2026 grafana.com |
| Session replay | Yes rrweb, browser DOM only our docs | Partial Add-on in public preview grafana.com |
| Browser click linked to its backend trace | Yes x-obs-interaction links a click to its trace, logs, and replay (browser SDK or header) our docs | Yes Faro user actions with related traces grafana.com |
| CPU profile from a trace | Yes When profiling is enabled and samples carry the trace ID our docs | Yes Span profiles for Go, Java, Ruby, .NET, and Python grafana.com |
| One ID chain from user to agent action | Yes user_id → session_id → interaction_id → trace_id → span_id → action_id our docs | Yes Frontend sessions link to backend traces; AI traces span user request to LLM grafana.com |
| Write your own queries over telemetry | Partial Search and filters; no query language or SQL editor our docs | Yes TraceQL, LogQL and PromQL for traces, logs and metrics grafana.com |
| Points out gaps in your instrumentation | Yes Flags spans with unexplained self-time in the waterfall and in get_trace our docs | Yes Application Observability detects uninstrumented services from trace data grafana.com |
Deployment and access
| Capability | Observability Unified | Grafana Cloud |
|---|---|---|
| Run the full platform yourself | Yes Docker image, Node + Postgres/S3, or Cloudflare our docs | Yes Self-managed open source or Enterprise stack grafana.com |
| Vendor-hosted cloud service | No Self-hosted only our docs | Yes Grafana Cloud grafana.com |
| Built-in single sign-on | No Front it with Cloudflare Access or oauth2-proxy our docs | Yes OIDC in all editions; SAML in Enterprise and Cloud grafana.com |
Choosing between them
Choose Grafana Cloud if you want a composable stack you can scale per signal, hosted or self-managed.
Choose Observability Unified if you want AI agents and engineers to investigate the same production evidence, with the telemetry kept on infrastructure you run. It is early-stage software suited to single-team deployments; the documented trade-offs include no built-in SSO, OTLP over HTTP only, and browser-only session replay.
How these cells were checked
Yes and Partial cells cite the vendor's own docs, pricing, product page, or official repository. No is used only where the cited page says so. Not marketed† means we read the vendor's product pages and found no such capability; it is an inference, not a vendor statement. Not stated means the vendor's docs do not say either way.
Rows for MCP write defaults, MCP links, agent tool calls, CPU profiles, self-hosting, and SSO were read on 2026-11-10 (ledger with excerpts). Rows for release and flag comparison, proof of cause, fix verification, gated remediation, click to trace, and dependency contention were read on 2026-11-11 (ledger with excerpts). Rows for compacted evidence for agents, agent skills, one ID chain, and instrumentation gaps were read on 2026-11-11 (ledger with excerpts). Rows for tool-call side effects, MCP call recording, production traces to eval cases, incident memory, and ad-hoc queries were read on 2026-11-11 (ledger with excerpts). The MCP server row was read on 2026-11-10 (ledger), and the remaining rows on 2026-10-13 (ledger). Vendors change their products often; if a cell is out of date, send a correction. The full comparison covers all nine vendors with excerpts.