Sea of Fox RTS MCP
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Sea of Fox RTS MCP develop the home island, train a landing force, and take the nearest objective"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Sea of Fox RTS MCP
Let your own agent play 狐狸之海:群岛战争 through five MCP tools. The adapter runs on your machine and connects directly to a compatible Sea of Fox RTS server. Your browser takes command when you enter the room. When you leave the room, a connected agent regains command of the same units, treasury, and orders.
Pair and connect
Enter a Sea of Fox RTS room with an updated game and server. For PvE or PvP, start the match in the game first.
Open 设置 → RTS MCP → 生成凭证 and copy the server address and credential. You can rotate or revoke this credential from the same menu.
Use Node.js 22.23.4 or later. In an MCP host that supports local stdio servers, configure the following. Put the credential in the host’s private environment configuration.
{
"mcpServers": {
"sea-of-fox-rts": {
"command": "npx",
"args": ["-y", "sea-of-fox-rts-mcp@0.1.1"],
"env": {
"SEA_OF_FOX_RTS_URL": "https://your-game-server.example",
"SEA_OF_FOX_RTS_TOKEN": "paste-your-mcp-credential"
}
}
}
}To try a development checkout, run npm ci and use node with the absolute path to index.mjs as the host's command and argument. For a local test game, set the RTS server address to http://127.0.0.1:5192. Use HTTPS for remote servers.
If the MCP API responds with HTTP 404, check the configured server address and whether the server operator has deployed MCP support. If your credential expires or is revoked, generate a new one in the game and update SEA_OF_FOX_RTS_TOKEN.
Ask your agent to inspect the match while you're connected. Then leave the game room and ask it to run a specific objective, such as “develop the home island, train a landing force, and take the nearest objective.” Keep your MCP host and agent task running if you want play to continue while the browser is closed.
Related MCP server: sc-bwapi-mcp
Tools
Tool | What it returns or does |
| Current faction rules, prices, unit roles and stats, training queues, resource production rates |
| Control status, match objectives, treasury, visible units and buildings, islands, resource reserves, recent events |
| Candidate construction sites on the builder's island, including dock locations |
| Up to 12 typed orders with individual acknowledgements; includes economy, construction, training, formations, naval transport, combat, and capture |
| Wait up to 20 seconds for control, a tick, a phase, a visible event, construction, a voyage, or a unit state |
See the play guide for a practical opening and protocol notes for control and delivery semantics.
Develop
npm ci
npm test
npm pack --dry-runTo test with a checkout of the game and installed RTS server dependencies, set SEA_OF_FOX_GAME_DIR to that checkout and run npm test. The integration test starts its own local room and talks to the CLI over MCP stdio. See the release process.
Available Tools
5 toolsrts_build_sitesBRead-only
Find server-validated candidate coordinates for a building on the land builder’s current island. Candidates can become invalid as the match changes. For docks the game searches the shoreline.
Inputs: kind (required), builderId (required), near.
rts_catalogBRead-only
Read the current faction rules, costs, building train lists and per-island resource production. Resources are generated by finished production buildings.
Inputs: no named parameters.
rts_observeARead-only
Read the current match, controlling peer, treasury, economy, units, visible contacts, or terrain landmarks. Use overview first; use paginated units/buildings/islands/reserves for IDs and positions. Events are buffered briefly.
Inputs: id, kind, view, limit, +4 more.
rts_orderADestructive
Apply up to 12 orders in sequence. The server validates control, ownership, placement, fog, resources and routes. Orders can partially succeed. Stop on the first rejection by default. Supply a fresh batchId and reuse it with exactly the same orders to recover an uncertain acknowledgement.
Inputs: orders (required), batchId, continueOnError.
rts_waitARead-only
Wait for an observed condition for up to 20 seconds, then return a fresh overview. Timeout is a normal result with matched=false. For a voyage the wait matches arrival or a blocked route; inspect the ship detail before landing. For landing use type=unit, status=ashore and the islandId.
Inputs: until (required), afterTick, timeoutMs.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
Each tool occupies a clearly separate lane: rts_order writes actions, rts_observe reads dynamic match state, rts_catalog reads static faction/economy rules, rts_build_sites answers a narrow placement query, and rts_wait blocks on a condition. The only mild adjacency is that rts_wait returns a fresh overview like rts_observe, but its condition/timeout semantics make the boundary obvious.
All tools share a uniform rts_ prefix with short, snake_case suffixes, so the set reads as one family. The suffixes mix verb-style (observe, wait, order) with noun-style (catalog, build_sites), which is a minor stylistic deviation rather than a real inconsistency.
Five tools is well-scoped for an agent-facing RTS interface: each tool absorbs a coherent, multiplexed capability (batched orders, multi-mode observation) instead of exploding into dozens of narrow calls. No tool looks vestigial, and none is doing two unrelated jobs.
The surface covers the core loop: discover rules/costs, observe state, locate legal build sites, issue validated orders, and await outcomes, with an idempotent batchId recovery path. Gaps are minor (e.g. no explicit tool for resigning, chat/diplomacy, or inspecting rejected-order details beyond the order response), and agents can work around them.
Maintenance
Related MCP Connectors
Make games with AI: one MCP server that connects your agent to Unity, Unreal and Godot editors.
A persistent multi-agent world any AI agent can join over MCP, with a public record of every match.
The game engine built for AI. Any agent makes a multiplayer game and publishes it.
Trading MCP server for AI agents, with live market data, account reads and controlled execution.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to read and control Command: Modern Operations scenarios locally via MCP tools, supporting real-time battle assessment, task planning, and scenario editing.11MIT
- FlicenseAqualityDmaintenanceEnables AI assistants such as Trae or Claude to control StarCraft 1.16.1 matches through BWAPI, providing tools to read game state, train units, build, move, attack, gather, and chat.9-
- FlicenseNot gradedqualityCmaintenanceEnables AI agents to connect to and control a windowed copy of Romance of the Three Kingdoms IX with Power Up Kit, letting them observe the game UI, gather officer/city intelligence, issue facility/transport commands, and advance turns through exposed MCP tools.-
- AlicenseNot gradedqualityBmaintenanceEnables an LLM agent to play a melee match against the game's built-in AI by exposing raw wc3env orders as MCP tools, with games graded and native replays saved.Apache 2.0