@calca/mcp-core

Pure, runtime-neutral half of Calca's Layer-2 agent playground: canvas tool implementations over @calca/canvas-base, protocol types, and the single-source tool contract that generates MCP tool JSON Schemas and docs.

No Node APIs, no MCP SDK wiring — the calca CLI (@calca/mcp) binds this surface to stdio transports.

Surface

| Tool | Effect | | ---------------- | ------------------------------------------------------------- | | canvas.read | summary / viewport / selection / full reads | | canvas.propose | Default mutation mode — records a proposal, never mutates | | canvas.apply | Lands commands (or a proposal) with origin: 'agent' | | canvas.lease | Exclusive soft leases, ttl clamped to 30 s | | canvas.release | Early lease release | | canvas.comment | type: 'comment' shape record pointing at a target | | canvas.arrange | row / column / grid layout via props.position updates |

import { createCanvasSession, callTool } from "@calca/mcp-core";

const session = createCanvasSession({ agentId: "claude-code" });
await callTool(session, "canvas.apply", { commands, requestId: "req-1" });

Semantics: every mutation returns {applied, commands, affected, requestId}; errors are {error: {code, message}} (duplicate | leased | op-limit | invalid | not-found). Request IDs are idempotency keys — a replay is a no-op. Ops are bounded to MAX_OPS_PER_CALL (100) per call.

Scaffolds pinned for later phases: src/wire (Phase 3, #41), src/crypto envelope crypto (Phase 4, #42).