Skip to main content
Manowar gives each agent a small default tool surface, then lets the agent discover heavier capabilities only when needed. The default sentence injected into the prompt is intentionally short:
That sentence is backed by real tools, not prose instructions. The runtime creates JSON-schema function tools, normalizes card-provided MCP and onchain actions, compacts large tool results, retries retryable failures, and enforces cloud permission gates when cloud mode requires them.

​
Default Surface

Card tools are loaded from the agent’s plugin bindings. If an agent card declares onchain or MCP plugins, Manowar fetches those tool schemas and exposes them directly when there are only a few actions. When there are many actions, it wraps them behind a card gate so the prompt sees one compact selector instead of a wall of functions.

​
Tool Flow

​
Permission Gates

Cloud mode can require explicit permission for sensitive actions. Manowar infers common consent categories from tool names, descriptions, and errors: When a permission is missing, the tool returns a structured CONSENT_REQUIRED error instead of silently running.

​
Result Compaction

Tool outputs are compacted before they go back into the agent loop. Defaults come from AGENT_TOOL_RESULT_MAX_CHARS, AGENT_TOOL_RESULT_ARRAY_ITEMS, AGENT_TOOL_RESULT_OBJECT_KEYS, AGENT_TOOL_RESULT_STRING_CHARS, and AGENT_TOOL_RESULT_MAX_DEPTH. This keeps tool use practical in long runs. The agent gets the useful part of a result, plus explicit truncation metadata when a result is too large.