mqttable_runtime_infoReads configured and effective MCP runtime status without exposing credentials.
Mqttable for AI Agents
Give AI agents MQTT evidence, not uncontrolled access.
Mqttable lets compatible agents discover the runtime, read bounded Trace and product evidence, create masked plans, and verify outcomes through a compact MCP surface.
Synthetic walkthrough: static text, not live telemetry. Mqttable does not call or bundle an independent model.
mqttable · mcp
Static transcript01 · Discover
$ mqttable_runtime_info
search_mqttable_capabilities → get_mqttable_schemaRuntime is loopback, effective access is approved_changes, and the schema includes its compact example.
02 · Capture
$ execute_mqttable_read → mqtt.qos.analyze_flow
execute_mqttable_read → proxy.toxic.listQoS analysis finds repeated PUBLISH packets and a PUBACK at 2.6 s; Proxy reports 2400 ms downstream latency.
03 · Reproduce
$ read_mqttable_resource → bounded Trace
execute_mqttable_read → proxy.toxic.listBounded Trace, Proxy, and Monitor evidence align on downstream latency while local runtime pressure stays normal.
04 · Act
$ plan_mqttable_action({ action: "proxy.toxic.delete", … })
apply_mqttable_plan({ plan_id, verification: "auto" })A durable masked plan targets only latency_downstream; apply accepts plan_id plus verification, not a replacement payload.
05 · Verify
$ verify_mqttable_plan({ plan_id })
execute_mqttable_read → mqtt.qos.analyze_flowVerification rereads state: the toxic is absent and PUBACK returns at 180 ms, inside the retry window.
QoS 1 latency retry
Discover, capture, reproduce, act, and verify against one bounded incident.
The transcript follows the current compact MCP contract. Read tools remain no-effect; changes use a durable masked plan and only apply when this local runtime is effectively in approved_changes.
- 01
Discover
Runtime is loopback, effective access is approved_changes, and the schema includes its compact example.
- 02
Capture
QoS analysis finds repeated PUBLISH packets and a PUBACK at 2.6 s; Proxy reports 2400 ms downstream latency.
- 03
Reproduce
Bounded Trace, Proxy, and Monitor evidence align on downstream latency while local runtime pressure stays normal.
- 04
Act
A durable masked plan targets only latency_downstream; apply accepts plan_id plus verification, not a replacement payload.
- 05
Verify
Verification rereads state: the toxic is absent and PUBACK returns at 180 ms, inside the retry window.
Compact external contract
Nine tools keep discovery narrow and behavior explicit.
Examples are included in schema expansion. Capability depth can change without teaching agents a new top-level tool list.
mqttable_doctorReturns passive diagnostics for runtime, transport, and discovery problems.
search_mqttable_capabilitiesSearches actions, workflows, resources, and prompts only when the capability id is unknown.
get_mqttable_schemaReturns one action, resource, or workflow schema with an inline compact example.
execute_mqttable_readExecutes one typed no-effect action through the compact read envelope.
read_mqttable_resourceReads one bounded mqttable:// resource through the tool facade.
plan_mqttable_actionCreates one immutable durable plan for an exact mutable action and returns a masked diff.
apply_mqttable_planApplies a stored plan using only plan_id and verification when the local change gate is effective.
verify_mqttable_planRereads runtime or persisted state for the same plan and returns verification evidence.
Persistent local policy
Access is a Settings choice, not a prompt on every action.
The user selects one of three modes in Mqttable Settings. An agent cannot elevate that mode through MCP.
offMCP unavailable
The endpoint is off. URL-only client configs work again after the user re-enables access.
read_onlyInspect without mutation
Trusted local clients can discover capabilities and read bounded evidence, but apply remains disabled.
approved_changesAllow plan-based changes
Trusted local clients may apply durable masked plans under the persisted local policy.
If readiness, registry, generation, or another mutation gate is not healthy, effective access automatically falls back to read_only even when approved_changes remains configured.
Hard boundaries
Agent access stays narrower than the workbench.
Writes stay on the local machine
Local MCP clients connect to the loopback endpoint without a client credential. Browser Origin requests are rejected, and mutable calls still require the live change gate.
Remote MCP is not included
MCP accepts loopback HTTP only. Non-loopback requests have no configuration or authorization-provider bypass.
Raw secrets stay outside MCP
Passwords, bearer tokens, PEM certificates, and private keys are entered through first-party intake. Agent flows reference SecretRef, CertRef, and KeyRef handles.
Replay stays in Workbench
Replay scenarios are created, reviewed, dry-run, and executed in the Workbench. There is no current Replay-specific MCP execution surface.
- 01
Open Settings → MCP, enable read-only, and choose whether this trusted local client may apply approved changes.
- 02
Copy the generated URL-only client configuration. MCP off/on keeps it valid without a client credential lifecycle.
- 03
Read runtime info first, search only when the capability id is unknown, then read its schema before calling it.
Agent FAQ
The useful questions before connecting an agent.
Give your agent bounded MQTT evidence without giving up control.
Open the Download page, choose the MCP access mode in Settings, and start with runtime discovery rather than guessed state.