@buzzr/mcp

@buzzr/mcp

MCP server exposing the @buzzr sports engines to AI agents — DFS settlement, odds math, and entertainment predictions as tools.

@buzzr/mcp is a Model Context Protocol stdio server that puts the whole @buzzr engine family in front of any MCP-capable agent (Claude Desktop, Claude Code, or your own client). All math and policy logic lives in the underlying engines — this package is a thin, schema-validated tool surface:

• @buzzr/dfs-engine — pick-em settlement with versioned compatibility and custom policies
• @buzzr/bets-core — no-vig fair lines, parlay pricing, expected value, Kelly staking
• @buzzr/entertainment-engine — ML buzz-score predictions and personalized game ranking

Install

Run it directly with npx (Node 22+). Pin the version you reviewed instead of silently accepting a future latest release:

npx -y @buzzr/mcp@5.1.0

The server speaks MCP over stdio: JSON-RPC on stdin/stdout, logs on stderr. It exposes 11 tools with bounded inputs and outputs, and does not fetch live odds, box scores, operator accounts, or private user data.

Claude Desktop

Add this to claude_desktop_config.json from Settings → Developer → Edit Config, then fully quit and reopen Claude Desktop:

{
  "mcpServers": {
    "buzzr": {
      "command": "npx",
      "args": ["-y", "@buzzr/mcp@5.1.0"]
    }
  }
}

Claude Code

claude mcp add --transport stdio buzzr -- npx -y @buzzr/mcp@5.1.0

For a version-controlled project configuration, add this to .mcp.json, trust the project when prompted, and start a new Claude Code session:

{
  "mcpServers": {
    "buzzr": {
      "command": "npx",
      "args": ["-y", "@buzzr/mcp@5.1.0"]
    }
  }
}

Confirm it with claude mcp get buzzr; inside Claude Code, /mcp shows the connection and discovered tool count.

Cursor

Add this to the repository's .cursor/mcp.json (or the equivalent user-level MCP configuration), then fully restart Cursor:

{
  "mcpServers": {
    "buzzr": {
      "command": "npx",
      "args": ["-y", "@buzzr/mcp@5.1.0"]
    }
  }
}

Open Cursor Settings → MCP to confirm buzzr connected and exposed 11 tools. See Cursor's MCP documentation if your installed Cursor version presents a different settings location.

Codex

The Codex CLI, IDE extension, and app share config.toml. The one-command user setup is:

codex mcp add buzzr -- npx -y @buzzr/mcp@5.1.0

Or add the equivalent block to ~/.codex/config.toml for all projects, or to a trusted repository's .codex/config.toml for that project only:

[mcp_servers.buzzr]
command = "npx"
args = ["-y", "@buzzr/mcp@5.1.0"]

Run codex mcp get buzzr, then start a new Codex task after changing the configuration.

Generic stdio MCP client

Use this process configuration in any client that accepts a command plus args:

{
  "name": "buzzr",
  "transport": "stdio",
  "command": "npx",
  "args": ["-y", "@buzzr/mcp@5.1.0"],
  "env": {}
}

Treat stdout as JSON-RPC only, read diagnostics from stderr, and close the child process's stdin during shutdown. The client must perform the MCP lifecycle below; starting npx in a terminal and seeing it remain open is normal for a stdio server waiting for a client.

Initialization and capability discovery

The client sends initialize first. Verify the serverInfo.name field is "buzzr", the serverInfo.version field is the installed package version, and the response has a capabilities.tools object. Send notifications/initialized, then call tools/list. The server returns the 11 tools in the catalog below with their input schemas. It does not advertise data fetching, resources, or prompts.

examples/mcp-calls.json is a machine-readable JSON-RPC transcript with the initialization expectation, exact tool discovery order, and a safe list_book_policies → validate_dfs_entry → grade_dfs_entry workflow. The repository replays that workflow through a real MCP client in CI.

Startup troubleshooting

• Confirm Node.js 22+ and npm are visible to the same desktop process or shell: node --version, npm --version, and npx --version.
• Confirm the pinned release exists with npm view @buzzr/mcp@5.1.0 version, then run npm cache verify. If npm reports cache corruption, repair npm's cache before retrying; npm cache clean --force is a last resort because it removes the whole local cache.
• GUI apps may inherit a smaller PATH than an interactive shell. On macOS or Linux, run command -v npx; on Windows, run where.exe npx. If the client cannot find npx, use that absolute path as command (normally npx.cmd on Windows) or launch the client from an environment where Node is on PATH.
• Keep logs off stdout. A healthy manual start writes a listening on stdio diagnostic to stderr and waits for JSON-RPC input.
• After any config, Node, PATH, or package-version change, fully quit and restart Claude Desktop, Claude Code, Cursor, or Codex. Opening only another chat tab may leave the old MCP child process and cached tool catalog in place.

Tool catalog

Tool	Engine	What it does
grade_dfs_entry	@buzzr/dfs-engine	Settle one transport-bounded 1–12-leg entry; executable built-in policies currently allow at most 6 PrizePicks or 8 Underdog legs.
grade_dfs_entries	@buzzr/dfs-engine	Settle 1–50 entries, up to 600 total legs, with bounded concurrency and isolated failures.
validate_dfs_entry	@buzzr/dfs-engine	Return structured engine validation issues for a candidate entry without settling it.
list_book_policies	@buzzr/dfs-engine	List authoritative executable profile snapshots and metadata-only drafts, including status, verification, sources, and complete play types.
fair_line	@buzzr/bets-core	Remove vig from both sides of one two-way market.
closing_line_value	@buzzr/bets-core	Compare placed and closing prices for the same selection.
parlay_value	@buzzr/bets-core	Price independent parlay legs, compare offered odds, and optionally calculate expected value.
kelly_stake	@buzzr/bets-core	Calculate full and fractional Kelly stakes from a supplied win probability.
summarize_bet_history	@buzzr/bets-core	Summarize up to 500 bets with overall and UTC-period rollups, drawdown, and streaks.
predict_game_buzz	@buzzr/entertainment-engine	Predict one game's 1–10 entertainment score with confidence and factor detail.
rank_games	@buzzr/entertainment-engine	Rank 1–100 candidate games for a bounded taste profile.

Tool strings, identifiers, arrays, stdio frames, concurrent calls, and serialized results are bounded. American odds must be within [-100000, -100] or [100, 100000]. grade_dfs_entries, closing_line_value, and summarize_bet_history return string contractVersion: "1".

DFS policy safety

Operator-named policies are independent compatibility profiles, not official rules engines or evidence of affiliation:

• PrizePicks is experimental and partially verified. Standard payout references were reviewed on 2026-07-16 from PrizePicks Payouts and PrizePicks Potential Outcomes, with DNP behavior checked against DNPs, Reboots, and Ties. The compatibility policy treats a 2-pick Power entry with a DNP as a refund when it falls below the two-pick minimum. Other settlement behavior and variable lineup-specific payouts remain incomplete.
• Underdog is experimental and unverified. The Underdog Sports Legal Center is the recorded rules entrypoint; the current compatibility payout and settlement values have not been verified.

Displayed lineup terms are authoritative. Call list_book_policies before grading, inspect verification and sources, and obtain explicit operator rulings for DNPs, reboots, ties, rescues, voids, and corrections. The tool lists future fixtures with executable: false; grading tools reject draft book IDs because drafts are metadata, not settlement implementations.

Error contracts

• Invalid tool arguments return the same bounded invalid_input result through an MCP client transport or a direct exported tool.handler(...) call. At most eight compact validation issues are included; raw Zod errors are never returned.
• Malformed JSON-RPC envelopes remain protocol errors owned by the MCP SDK and are distinct from a valid tools/call request with invalid tool arguments.
• validate_dfs_entry intentionally accepts a bounded candidate object and returns the engine's structured validation report.
• Execution failures return generic isError: true results such as tool_execution_failed, entry_settlement_failed, server_busy, or result_too_large. Internal error details are not public.

Example transcripts

"Did my PrizePicks slip hit?"

User: I had a $10 2-pick power play: LeBron over 25.5 points (he scored 31) and Curry over 27.5 (he scored 33). What's the payout?

Agent calls grade_dfs_entry:

{
  "entryId": "slip-1",
  "bookId": "prizepicks",
  "playTypeId": "power",
  "stake": 10,
  "displayedMultiplier": 3,
  "legs": [
    { "legId": "l1", "playerName": "LeBron James", "league": "NBA", "propType": "points", "line": 25.5, "direction": "over", "actual": 31 },
    { "legId": "l2", "playerName": "Stephen Curry", "league": "NBA", "propType": "points", "line": 27.5, "direction": "over", "actual": 33 }
  ]
}

Result: "status": "won", "payout": { "total": 30, "withdrawable": 30, "bonus": 0 } — both supplied actuals clear their lines and the submitted 3× displayed multiplier is consistent with the selected compatibility table. Confirm the actual entry details before treating this as an operator outcome.

"Is this parlay +EV?"

User: The book offers +264 on a two-leg parlay where both legs are -110/-110 markets. Worth $100?

Agent calls parlay_value with { "legs": [{ "selected": -110, "opposite": -110 }, { "selected": -110, "opposite": -110 }], "offeredAmericanOdds": 264, "stake": 100 }.

Result: fair probability 0.25 (fair price +300) vs. offered +264 — "edgePercent": -2.47 and an expected loss of about $9 per $100. The agent recommends passing.

"What should I watch tonight?"

User: I'm a Lakers fan. Rank tonight's slate for me.

Agent calls rank_games with the slate and { "profile": { "favoriteTeams": ["Lakers"], "leagueAffinity": { "NBA": 0.8 } } }.

Result: each game comes back with baseScore, bounded affinityAdjustment and socialAdjustment, a totalScore, and named factors ("favorite team playing", "league affinity") the agent can cite when explaining the ranking.

Embedding

The server is also exported as a library, so you can mount the same tool catalog on your own transport (in-memory for tests, HTTP, etc.):

import { createBuzzrMcpServer, allTools } from '@buzzr/mcp';

const server = createBuzzrMcpServer();
await server.connect(myTransport);

Individual tool definitions (gradeDfsEntryTool, fairLineTool, …) are exported too — each is { name, title, description, inputSchema, handler }, and handlers can be called directly without any transport.

Codex skill

The repository includes a Buzzr Sports Engine skill with the 11-tool routing guide, limits, response-reading order, and operator-safety rules:

npx skills@1.5.17 add https://github.com/Buzzr-app/dfs-engine --skill buzzr-sports-engine --agent codex --yes --copy

Repository contributors can prove the same one-command install without using a published branch: npx skills@1.5.17 add . --skill buzzr-sports-engine --agent codex --yes --copy. npm run check:skill runs the official pinned quick_validate.py, performs that local install in an isolated temporary project, and verifies every installed skill file byte-for-byte.

Compatibility

• Node.js >= 22 is supported.
• rank_games requires @buzzr/entertainment-engine >= 5.0.0. Against an older engine build the tool degrades gracefully with an engine_capability_missing error result instead of crashing the server.

Import the supported API from @buzzr/mcp. Deep src/* and dist/* imports are unsupported. See the all-package API index and the generated root-export reference.

Report reproducible defects in GitHub Issues. Report vulnerabilities privately through SECURITY.md. The versioning and support policy defines the supported runtime and SemVer contract.

License

MIT