MCP
BattleGrid can talk to compatible clients outside the product UI — the same core actions you already use through Commander: strategies, agents, market reads, Radar, Arena, and trade control.
For most players: start in the UI or Commander. Connect MCP only when you want those same verbs from Cursor or another automation client. There is no second, secret API — your client sees what your permissions allow.
How it fits
| You want… | Use |
|---|---|
| Fast plain-language control in the product | Commander |
| Prompt patterns that work well | Prompt formula |
| The same actions from Cursor / other MCP clients | MCP (this page) |
Commander discovers this catalogue at runtime and family-gates it. An MCP client sees the tools your OAuth scopes allow.
What players should know
- Read vs spend. Research, library, and non-financial setup stay on the lighter permission. Anything that moves money, submits a wager, or grants standing autonomous authority (Radar deploy, Arena deploy, halt/resume that changes whether capital can commit, close position, accept a pending entry, submit a grid) needs the stronger wager permission — and usually an explicit confirm.
- Radar vs Arena. Creating or replacing a Radar deployment needs the wager permission because a policy is standing authority to trade that coin. Creating or replacing an Arena deployment needs it because a live policy can auto-submit grids and charge entry fees. Arena slots still configure game participation only — they do not grant continuous trade permission by themselves. To trade a coin, deploy on Radar.
- Strategy and agent changes. Every change is staged into your draft — the same draft the Strategy Builder and the agent console show you — and nothing changes live until it is committed, on your word, naming the draft version and revision your client read. If either moved in the meantime, the commit is refused and the client reads the draft again. Same path as Commander. See Strategies.
- Safety habits. Prefer reads until the ask is clear. On a client disconnect mid-write, read back state before retrying — blind retry can double a write. Use Prompt formula the same way you would in Commander.
For automation / integrators — scopes, revisions, tool names
Scopes
Two scopes. A family never straddles them.
| Scope | Label (consent) | What it allows |
|---|---|---|
mcp:read | Account & configuration access | Account and market information; create / update / bind / archive / restore non-financial configuration. Cannot submit wagers or move funds. |
mcp:wager | Wager / standing authority | Money movement or standing autonomous authority — Arena/grid submits, Radar deploy upsert/delete (including pause via enabled), agent halt/resume and related runtime wager tools, close position, accept/cancel pending entry decisions, and similar. |
Grant mcp:wager only to clients you trust. Destructive deployment and trade tools additionally require an explicit confirm: true (or the tool’s documented confirm field); a missing confirm is a refusal, not a soft no-op. Agent and strategy commits take no confirm: the draft read shows what a commit will do, and your client asks you before committing.
Capability map (by intent)
Grouped the way Commander skills think — not an exhaustive tool dump.
| Intent | Read side (examples) | Wager / authority side |
|---|---|---|
| Account | Balance, slots, game stats | — |
| Market / coins | Regime, candles, rankings, coin qualification scans | — |
| Strategies | Library, templates, report preview; stage, read and commit the strategy's draft (configuration) | — |
| Agents | List, performance, budget, qualification, activity; stage, read and commit the agent's draft — create, edit, rebind (config); archive (config) | halt / resume and related runtime (wager) |
| Radar | List/get deployments, preview resolution, activity summary | Upsert / delete deployment (wager) |
| Arena / grids | Presets, policies, preview resolution, generate grid | Upsert/delete deployment policy; submit grid (wager) |
| Live trades | Open positions, entry decisions, charts | Accept/cancel entry; close position; protection overrides (wager) |
A stage writes your draft, not the live agent or strategy. A commit publishes exactly the draft version and over exactly the revision your client read, and a retry of the same commit returns the same result. Same path as Commander.
Authoring contract (agents and strategies)
- Discover vocabulary / section templates as needed.
stage_strategy_draft(orstage_agent_draft) with{ strategyId?, draftVersion, axes }— omit the id to start a new one. Stage reports problems in the draft; it does not refuse them.get_strategy_draft(orget_agent_draft) — the draft, itsdraftVersion, the livecommittedRevision, a per-axis diff, diagnostics and impact (blast radius). Human (or your agent) reviews it.- On the player's word,
commit_strategy_draft(orcommit_agent_draft) with exactly thatdraftVersionandexpectedRevision(nullfor a new one). A refusal names the next act:nextAct: "get_draft"(read the draft again) ornextAct: "stage"(fix what the diagnostics named).
Revision fields (draftVersion, expectedRevision, sourceRevision, and related) are optimistic concurrency. Stale values conflict; nothing is overwritten silently. On the player surfaces this reads as “save refuses stale edits.”
Radar and Arena via MCP
- Radar — coin-scoped policy; every matching rule's agent on duty, each rule reading the confirmed regime at its agent's regime timeframe; rule/default slots, the higher rule taking a shared close. Preview is dry-run and reports one row per on-duty agent. Upsert/delete are wager-scoped. Field guide: Radar.
- Arena — game-preset deployment; slots configure game participation only and do not grant trade permission by themselves. To trade a coin, deploy on Radar. Upsert/delete are wager-scoped.
Connecting a client
Authorize an MCP client against BattleGrid with the scopes you intend (mcp:read alone for configuration and research; add mcp:wager only when the client must deploy capital or standing authority). Live discovery is the source of truth for parameter shapes on strategy studio tools.