stroblme
23bb9aa983
Every output is a target and every input besides prompt a reading. A targets setting gives descriptions, values, ranges and per-target guards, model takes a trained pack's artifact digest, min_confidence defaults to the model's threshold. A command is applied whole or not at all, and targets the model was not trained on read degraded. The home path (area/kind ports, groups, brightness, temperature answers) is gone. Claude-Session: https://claude.ai/code/session_01H89mFby73wwhthP9htfcuB
3.8 KiB
Agents over MCP
Fluksio exposes flow tools over MCP. An agent can inspect, edit, publish, and run flows with the permissions of the user who authorizes it.
Connect an agent
Set MCP_ENABLED=true in the engine's environment and restart it, under
environment in compose.yml for the Docker container. It defaults to off.
Configure your MCP client with an HTTP endpoint:
- Docker container:
http://localhost:8080/mcp - Deployment:
https://<your-domain>/mcp
The endpoint has to be the address in FRONTEND_HOST, which is also the
issuer of the client's tokens.
The client registers, opens browser authorization, and receives an OAuth
token after you approve access. The issuer and browser redirects use
FRONTEND_HOST and ENVIRONMENT, so configure those for the addresses your
client and browser can reach.
Manage registered clients under Admin → Agents. Revoking a client
invalidates its access. MCP uses a separate signing key at
OAUTH_PRIVATE_KEY_FILE; rotating that key invalidates all MCP tokens.
An instance connected to a portal also answers
through it, at https://hub.fluksio.com/i/<instance-id>/mcp (the portal's
own hub. host for a custom one). The portal forwards the endpoint and its
OAuth, and the instance checks the agent's token as it does locally. The
consent page opens in the instance as the portal serves it, so open the
instance from the portal first if this browser has not recently.
Available tools
| Task | Tools |
|---|---|
| Inspect flows | list_flows, get_flow, get_node_source, list_node_types, get_graph, list_shared_nodes |
| Inspect values | get_flow_state, get_message_history |
| Inspect the instance | list_secrets (names only), get_modules, get_health, get_metrics, list_failures, list_runs (live cascades) |
| Inspect devices | list_devices, get_device |
| Edit flows | save_flow, save_node_source, validate_flow, publish_flow, discard_draft, delete_flow |
| Operate flows | run_flow, trigger_node, start_flow, stop_flow, pause_flow, resume_flow, cancel_node |
| Install node packages | apply_modules |
Tools call the same API as the browser. Drafts, validation, version checks, and superuser restrictions apply. Secrets cannot be read back.
A useful workflow: ask the agent to inspect the flow, save a draft, and validate it. Review the draft before asking it to publish. The agent has a publish tool, so make your intent explicit.
The assistant
A superuser configures an OpenAI-compatible provider under Settings → Assistant, with a base URL, API key, and model. This enables the dashboard's Assistant entry for signed-in users.
The assistant uses the same tools with your own permissions. Draft edits
appear in the editor. Each answer streams in as the model writes it and
renders as markdown with a tool trace.
Conversations are stored per user on the instance; the bin button clears
yours.
MCP_ENABLED is only needed for external clients, so the built-in assistant
works without it.
Control
A live Wren or
Assistant node with its prompt and a
reply or actions wired adds a Control tab to the chat. Wren uses its
local model, the Assistant node the provider above.
Choose a conversation and type a command. It goes out on the node's prompt,
which is where a stt node writes what a device heard, so typed and spoken
commands are one conversation: the words heard so far, the command, each
reply, and the reported actions with their target message, value and whether
each was applied. A node whose prompt is another's unhandled answers in that
node's conversation. Wren's apply, min_confidence and targets settings
decide whether it writes to its targets. Control conversations stay in the
browser and clear on reload.