MCP Surface

Bardo’s MCP surface is optimized for a trusted copilot, not a hidden autonomous game master. The tools are designed to stay discoverable, structured, and explicit about what is read-only, what is advisory, and what can affect canon.

Tool shape#

Every public tool is documented with a clear use case, a “when not to use” boundary, and structured output. The goal is to help clients route themselves correctly instead of improvising from vague descriptions.

The surface is tool-only in this phase. Bardo does not currently depend on MCP prompts or resources to expose the trusted-copilot GM workflow.

Core and mechanics tools#

The core surface covers canon retrieval, reporting, scene adjudication, and mechanics support:

• context_query
• world_state_overview
• timeline_diff
• player_knowledge_view
• continuity_audit
• scene_turn
• validate_action_against_ruleset
• resolve_mechanics
• roll_dice
• world_sync_preview
• ruleset_mechanics_overview

These tools are meant to compose cleanly. For example, an agent can inspect the current ruleset, validate a proposed action, resolve the mechanics when authoritative support exists, and fall back to scene_turn when broader canon-aware judgment is still needed.

Strategy tools#

Bardo also exposes planning-oriented tools so an agent can operate above the single-action level:

• faction_pressure_board
• campaign_clocks
• encounter_prep
• campaign_planning

These tools are intentionally read-only. They help an agent reason about pressure, pacing, and prep without pretending that planning output is already canonical truth.

Authoritative writes#

The main canon-affecting workflow still centers on scene_turn. resolve_mechanics can append authoritative mechanics events when a ruleset is fully supported, but raw mutation tools are not part of the public trusted-copilot surface.

That split is intentional: Bardo tries to give agents strong guidance, previews, and planning breadth without turning the public MCP surface into a low-level maintenance API.

Rate limiting and V1 guardrails#

The hosted surface is also guarded by request budgets. That matters most around bridge startup, status polling, and any route that can fan out into metering or approval checks.

For agents, the practical rule is simple: plan with the structured tools, call only the next best tool, and back off on 429 instead of retrying aggressively. The surface is designed for deliberate copilot behavior, not for brute-force probing.