MCP
Tools for agents that work safely in operations.
Approved command actions can be exposed as typed MCP tools. Tools follow the same approval boundary humans do. AI agents propose, policy decides, the receipt explains what happened.
What it does
Three things this surface gives you.
- The tool catalog stays aligned with approved command actions. New approved actions appear without manual doc work.
- Policy is uniform: a human, a system, or an agent calling the same tool gets the same checks.
- Run receipts include actor type, tool name, input, and policy decision for every call.
Hosted MCP endpoint
Worked example.
# Preview access provides KARMANFLOW_API_URL and a scoped agent key.
curl -s "$KARMANFLOW_API_URL/mcp" \
-H "content-type: application/json" \
-H "authorization: Bearer $KARMANFLOW_AGENT_KEY" \
-H "x-karmanflow-tenant: $KARMANFLOW_TENANT_ID" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq
# Agents call tools through the same registry the
# operator console reads. Risky tools route through
# RequestApproval before the underlying command runs.Continue inside the hosted KarmanFlow docs for the schema reference, explorer, runtime notes, and preview setup guidance.
Read hosted runtime guidePreview
Run a working integration path.
Request a guided preview workspace, or bring a sample integration shape to a working session and we will walk through it on a call.