Brute Genius
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., "@Brute Genius plan a full grid search over 5 hyperparameters, ranked by accuracy"
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.
Copyright © 2026 Lewis James Seals Jr. All rights reserved. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
Brute Genius
Exhaustive combinatorial search as an MCP server. Given a discrete search space, Brute Genius enumerates the full Cartesian product, shards it across a fleet of accelerators, scores every combination against a named objective, and returns a ranked list. It does not sample or prune; it tries everything.
Warning: this plugin spends real money fast. Always call brute_genius_plan first and respect the cost ceiling it returns. brute_genius_kill is the emergency stop.
Connect to Claude Desktop
Install dependencies:
cd brute-genius
pip install -e ".[test]"Add the server to Claude Desktop’s MCP configuration (
claude_desktop_config.json):
{
"mcpServers": {
"brute-genius": {
"command": "python",
"args": ["/absolute/path/to/brute-genius/server.py"]
}
}
}Restart Claude Desktop. The tools
brute_genius_plan,brute_genius_launch,brute_genius_status,brute_genius_results,brute_genius_kill, andxpu_probewill appear, along with the resources and prompts.
For other clients (Cursor, Copilot, etc.) point the MCP host at the same python /absolute/path/to/server.py stdio command. For remote/HTTP use:
mcp.run(transport="http", host="0.0.0.0", port=8000)Related MCP server: unmerged-approaches-mcp
Quick test
pytest test_server.py -v
npx @modelcontextprotocol/inspector python server.pySafety guarantees
plan()refuses any estimate abovemax_cost_usdHard caps: 64 nodes, 10⁹ combinations
Objectives are registry-only; caller code is never executed
kill()always stops every node associated with a plan_id
Available Tools
6 toolsbrute_genius_killBrute Genius KillB
Stop every node and shard belonging to a plan. Ends billing.
Partial results already scored are preserved. Always available even if the calling agent goes silent.
Inputs: plan_id (required). Returns: object
brute_genius_launchBrute Genius LaunchC
Shard the enumerated space and score every combination.
Re-checks balance and cost ceiling before any work is dispatched. Streams progress via ctx.report_progress and logs via ctx.info.
Inputs: plan_id (required), chip, nodes. Returns: object
brute_genius_planBrute Genius PlanA
Enumerate the full Cartesian product and return a cost estimate.
Does NOT execute anything. Refuses to return a plan_id when the estimated cost exceeds max_cost_usd.
Inputs: space (required), objective (required), max_cost_usd (required), chip, +1 more. Returns: object
brute_genius_resultsBrute Genius ResultsC
Return the top_k highest-scoring combinations (descending score).
Inputs: plan_id (required), top_k. Returns: result
brute_genius_statusBrute Genius StatusA
Return live progress, shards done, combinations evaluated, best-so-far.
Inputs: plan_id (required). Returns: object
xpu_probeXpu ProbeB
Detect and report available accelerators on the current machine/fleet.
Inputs: chip. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
The lifecycle tools are mostly clearly separated: probe, plan, launch, status, kill, and results each have distinct roles. The main possible confusion is between `brute_genius_status` reporting best-so-far and `brute_genius_results` returning top_k combinations, but the live versus final-results distinction is reasonably clear.
Five of six tools use a consistent `brute_genius_` prefix with readable action/resource suffixes. `xpu_probe` is the only outlier and follows a different prefixing convention, which is a minor deviation rather than a systemic inconsistency.
Six tools is well-scoped for a brute-force planning, execution, monitoring, and retrieval server. Each tool maps to a distinct phase of the workflow and there is no obvious bloat.
The tool set covers the full lifecycle: hardware probing, combinatorial planning with cost guards, launch/execution, live status, termination, and results retrieval. There are no obvious missing operations for the stated brute-force scoring domain.
Related MCP Connectors
Plan, run and inspect AI-agent drills with synthetic users, hosted tools and cloud worlds.
Commission a multi-model AI spec committee from your agent; get rubric-scored, build-ready specs.
Execution control plane for agents: capabilities, durable jobs, budgets, receipts, and BYOK.
Public read-only discovery of agent, model, training, task, and verification opportunities.
Related MCP Servers
- FlicenseAqualityBmaintenanceEnables coding agents to run multiple approaches or tasks in parallel on isolated Daytona sandboxes, returning tested diffs and benchmark rankings so the best outcome can be applied.4-
- AlicenseAqualityCmaintenanceEnables sending one identical brief to multiple LLMs in parallel and returning every answer verbatim, unmerged, unranked, and unsynthesized.411 npmMIT
- AlicenseAqualityBmaintenanceLets assistants send the same test prompts to as many as four text models across OpenRouter, Amazon Bedrock, Google Vertex AI and Microsoft Foundry, then automatically grade each answer with a judge model, rule checks and a policy gate. Returns scored leaderboards with letter grades, per-response strengths and weaknesses, latency, throughput and estimated cost, plus downloadable PDF and CSV reports.930 PyPIMIT
- AlicenseAqualityCmaintenanceEnables configuration-driven, budgeted multi-source web research across platform and web search adapters, merging duplicate links, preserving provenance, and returning honest source coverage reports.4MIT