MCP tool reference¶
Auto-generated by scanning
crates/mcp/src/tools.rs. Runscripts/docs/regen.pyto refresh.
AppControl ships an MCP (Model Context Protocol) server that exposes the production stack to any MCP-compatible client: Claude Desktop, Claude Code, Cursor, ChatGPT (via bridge), and others. The MCP server is a thin shim — every tool call resolves to one REST API call, with the same auth and audit semantics as a CLI or UI call. Every MCP-triggered action lands in action_log with actor_type='mcp'.
Setup¶
// ~/.config/claude-desktop/claude_desktop_config.json
{
"mcpServers": {
"appcontrol": {
"command": "appcontrol-mcp",
"args": ["--url", "https://appcontrol.example.com"],
"env": {
"APPCONTROL_API_KEY": "ac_xxxxxxxxxxxx"
}
}
}
}
The MCP binary is shipped alongside the backend; run appcontrol-mcp --help for the version-specific flags.
Tools¶
21 tools exposed.
| Tool | Required args | Description |
|---|---|---|
get_application_knowledge |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
propose_application_map |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
list_apps |
(none) | List every application the API key can see. |
get_app_status |
app_name |
Return current overall status, per-component states, and the FSM transition history for an application. |
start_app |
app_name, dry_run |
Start an application in DAG order. Pass dry_run=true to validate the plan without dispatching. |
stop_app |
app_name |
Stop an application in reverse-DAG order. |
diagnose_app |
app_name |
Run the 3-level diagnostic (Health / Integrity / Infrastructure) and return the recommendation matrix. |
rebuild_app |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
get_rebuild_status |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
get_incidents |
app_name, days |
Return incidents (FSM transitions to FAILED) over the last days days. Default: 7 days. |
get_topology |
app_name |
Return the DAG: nodes, edges (strong/weak), groups, and per-component metadata. |
estimate_time |
app_name, operation |
Estimate how long operation (start or stop) will take based on historical execution times for this application. |
get_activity |
app_name, limit |
Return the most recent rows from action_log for this application, capped by limit (default 20). |
list_agents |
(none) | List enrolled agents with hostname, last heartbeat, and version. |
list_log_sources |
app_name, component_name |
List the log streams an agent exposes for a given component (declared in log_sources on the component). |
get_component_logs |
app_name, component_name, source, lines |
Tail the last N lines from a component log stream. Useful for claude code-driven incident analysis. |
run_diagnostic_command |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
search_logs |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
file |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
event_log |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
command |
(none) | (undocumented — add to scripts/docs/gen_mcp.py::TOOL_DESC) |
Permissions and audit¶
MCP requests authenticate with the same API keys as the CLI and REST API:
- The key carries the user's effective permission.
start_appon an application where the key has onlyViewreturns the sameforbiddenerror as the REST endpoint would. - Every tool call is recorded in
action_logwithactor_type='mcp'and the tool name inaction_log.action. Audit reports surface MCP-triggered actions identically to UI-triggered ones. - Read-only tools (
list_apps,get_app_status, …) still appear in the audit trail — DORA Article 16 requires every access to be loggable, not just mutations.
Example session (Claude Desktop)¶
User: Which components of core-banking failed in the last 24h?
Claude: [calls get_incidents(app_name='core-banking', days=1)]
Two components transitioned to FAILED in the last 24h:
— batch-eod (07:14 UTC, exit_code=2)
— settlement-queue (11:02 UTC, exit_code=1, recovered at 11:08)
User: Show me the last 40 lines of batch-eod's stderr.
Claude: [calls get_component_logs(app_name='core-banking', component_name='batch-eod', source='stderr', lines=40)]