---
title: Prime Rush — a battle royale as an MCP environment
emoji: 🎯
colorFrom: blue
colorTo: yellow
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
short_description: A live battle royale as six MCP tools
tags:
  - agents
  - mcp
  - benchmark
  - games
  - tool-use
---

# Prime Rush Environment

> This Space holds the MCP adapter, the task pack and the reference grader,
> and runs an interactive demo. The benchmark — methodology, per-episode
> results and trajectories — is in the
> [dataset](https://huggingface.co/datasets/datagame/prime-rush-results).

> The demo drives a stand-in game server so it needs no infrastructure. For
> access to the real environment, get in touch.

A live battle-royale match from a shipped commercial game, exposed to an LLM
agent as six MCP tools. A safe circle shrinks on a timer and moves; the agent
is given its own coordinates and the circle's, and has to work out where to
go.

Built for the [Archipelago](https://github.com/Mercor-Intelligence/archipelago)
harness, following the conventions of the MCP servers already in that repo.

## Highlights

- **Deterministic stepping**: the match advances only inside a tool call, by
  the amount that call asks for. The same seed and the same calls reproduce
  the same match, and a slow model is not penalised.
- **No derived guidance**: `get_situation` returns raw coordinates. There is
  no `move_to(x, z)` and no bearing to the circle, because computing it is
  the task.
- **Snapshot grading**: every call that advances the match rewrites
  `outcome.json` from game state, so the score is read from a file rather
  than from the transcript.
- **Harness compatibility**: ships as an Archipelago task pack plus a
  reference eval definition.

## Tools

| Tool | Advances the match |
|---|---|
| `get_situation` | no |
| `turn_to_bearing` | yes, by the time the turn takes |
| `move` | yes, by the seconds requested |
| `hold_position` | yes |
| `shoot` | yes |
| `finish` | no |

## Results

Current scores are on the **Leaderboard** tab, read from the results dataset.
The methodology, confidence intervals and per-episode data are in the
[dataset card](https://huggingface.co/datasets/datagame/prime-rush-results).

## Running it

Neither test suite needs a game build; `tests/fake_game.py` stands in for the
server and reproduces its response characteristics.

```bash
pip install fastmcp pydantic
python tests/test_end_to_end.py      # 29 checks
python tests/test_grading_eval.py    # 9 checks
```

Against a real match:

```bash
export INDUS_BRIDGE_HOST=... INDUS_BRIDGE_PORT=...
python -m indus_mcp.main             # stdio, the transport Archipelago uses
```

Under Archipelago, `POST /apps` with `tasks/zone_survival/mcp_config.json`.

| Variable | Meaning | Default |
|---|---|---|
| `INDUS_BRIDGE_HOST` / `INDUS_BRIDGE_PORT` | the running match to attach to | `127.0.0.1:7850` |
| `INDUS_TASK` | task name | `zone_survival` |
| `INDUS_BUDGET_SECONDS` | game time the agent is given | `100` |
| `INDUS_FF_RADIUS` | hand over once the circle is this small | `350` |
| `INDUS_OUTCOME_PATH` | where the grader reads the result | `/filesystem/outcome.json` |
| `MCP_TRANSPORT` | `stdio` or `http` | `stdio` |

> `INDUS_FF_RADIUS` must match across every row of a comparison. Before the
> circle has closed it covers most of the map and any agent is inside it, so
> rows that handed over at different radii did not start in comparable
> situations.

## Limitations

- `shoot` is horizontal only. Aim pitch is not in the observation, so an
  agent cannot verify the effect of a `pitch` argument.
- No line of sight: nearby enemies are reported through walls.
- One task, one map.
- One agent per match. The game build supports multiple seats; the harness
  does not use them yet.

---

SuperGaming. Apache-2.0.