Bounty Operator

Bounty Operator MCP

Make your strongest case before you submit.

Check your proof, make the impact explicit, and catch missing evidence before the triager has to ask.

This is the MCP server for Bounty Operator. Your agent names the files, the server scans them for secrets and returns the review request, your agent's own model writes the review, and the server checks every cited line and builds the evidence packet. Preparing a review and building the packet need no account and no key.

That route covers the three core profiles. Every other profile, and seven of the eight gauntlet stages, runs as a hosted review on the Bounty Operator service with your own provider key.

Install

Two routes. Both work today.

Remote endpoint

Nothing to install. One line per client.

Claude Code

claude mcp add --transport http bounty-operator https://bountyoperator.com/api/mcp

Codex

codex mcp add bounty-operator --url https://bountyoperator.com/api/mcp

Cursor, in ~/.cursor/mcp.json or .cursor/mcp.json in a project:

{ "mcpServers": { "bounty-operator": { "url": "https://bountyoperator.com/api/mcp" } } }

Local server

This package. It reads files by path, keeps review preparation on your machine and adds run_gauntlet_plan. Node 22 or later, no dependencies. npx installs it from the tarball the site serves.

Claude Code

claude mcp add --transport stdio bounty-operator -- npx -y https://bountyoperator.com/dl/bounty-operator-mcp.tgz

Codex

codex mcp add bounty-operator -- npx -y https://bountyoperator.com/dl/bounty-operator-mcp.tgz

Cursor, in ~/.cursor/mcp.json or .cursor/mcp.json in a project:

{ "mcpServers": { "bounty-operator": { "command": "npx", "args": ["-y", "https://bountyoperator.com/dl/bounty-operator-mcp.tgz"] } } }

Any other client: command npx, arguments -y https://bountyoperator.com/dl/bounty-operator-mcp.tgz, transport stdio.

On Windows outside WSL, Claude Code starts npx through cmd: end the command with -- cmd /c npx -y https://bountyoperator.com/dl/bounty-operator-mcp.tgz.

The package is coming to npm as bounty-operator-mcp. Until it is published, the tarball URL above is the install.

A first request to try:

Use bounty-operator to prepare a Solidity review of src/Vault.sol. Answer the request it returns, then build the packet.

Tools

Six tools here, five on the remote endpoint: run_gauntlet_plan is local only.

Limits per call: 50 files, 120 KB per file, 240 KB and 20,000 lines in total.

Core and hosted profiles

prepare_review takes the three core profiles: general (Code security review), solidity (Solidity review) and report (Challenge a draft report). Their method is in this package and your agent's own model answers the request.

Every other profile is hosted: scope, provenance, prior-art, poc, severity, triage, report-edit and scanner. Each runs on the Bounty Operator service through run_review, on your own provider key. The method of a hosted profile is added on the service. It is not in this package and no tool returns it.

For a hosted profile:

• list_profiles returns it with hosted: true.
• prepare_review fails with the code hosted_profile before any file is read.
• run_review runs it and uses one hosted review. A review the provider blocks under its usage policy is not counted. The two profiles no list shows, verdict and panel, are the last step of the gauntlet and of a panel review: they run on Operator only.

{

"profile": "triage", "error": "Triager simulation runs on the server. Call run_review with profile \"triage\": it needs BOUNTY_OPERATOR_TOKEN and the provider key in this server's environment.", "code": "hosted_profile" }

Files by path

prepare_review and run_review take paths, files, or both.

• paths are read by the server from its working directory: ["report.md", "src/Vault.sol"]. The agent does not have to send the file text.
• files carry the text inline: [{ "name": "src/Vault.sol", "content": "..." }].

A path resolves under the working directory and nowhere else. The server refuses .., an absolute path outside the directory, and a link that leaves it. File contents go to the review and are never returned: a result holds the label, size, line count and SHA-256 of each file.

A client that starts the server outside your project sets BOUNTY_OPERATOR_ROOT to the project folder.

The privacy check

Every call scans the files first. A secret blocks the call and the result lists file, line and kind of each match, never the matched text. An email or IP address stops the call until you repeat it with acknowledgeWarnings: true. run_review runs this check on your machine before anything is sent.

Prompts

Three prompts, listed by clients that support them as slash commands:

In Claude Code: /mcp__bounty-operator__gauntlet.

The gauntlet

Eight stages, in the order that ends a weak report early: scope, provenance, prior art, proof, severity, triager, report, verdict. run_gauntlet_plan returns the plan and names the tool that runs each stage. Seven stages are hosted and run through run_review, so the run needs a connection token and a provider key. The report stage is a core profile: prepare_review returns its request and your agent's own model answers it. Each review is passed forward as stage- - .md, and the run ends with build_packet. The verdict is one of submit, rewrite-then-submit, prove-first, hold-duplicate, drop.

A run uses seven hosted reviews. Free covers one hosted review per UTC day, so a full run takes Operator. The plan carries the count under hosted, and account returns the allowance before the run starts.

Hosted reviews

run_review runs any profile on a provider's model with your own API key and returns the review checked. It is the only way to run a hosted profile from an agent. It uses the same allowance as the website: one hosted review per UTC day on Free, any profile, and unlimited on Operator at US$10 per month.

Three result fields say when the text is not a review. truncated is true when the provider cut the answer short. refused is true when the model or the provider declined. blocked is set when the review was blocked: anthropic-cyber, anthropic-reasoning or openai-cyber for that provider's safeguards, guardrail for a guardrail on the key or its account, policy for any other block under a usage policy. A blocked review is never counted. After a refusal or a block, do not call again with the same model: use another model or provider, or prepare_review for a core profile.

1. Sign in at bountyoperator.com and create a connection in the account panel. The token starts with bok_ and is shown once.
2. Start the server with the token and one provider key in its environment.

claude mcp add --transport stdio --env BOUNTY_OPERATOR_TOKEN=bok_... --env OPENROUTER_API_KEY=sk-or-... bounty-operator -- npx -y https://bountyoperator.com/dl/bounty-operator-mcp.tgz

Codex, in ~/.codex/config.toml:

[mcp_servers.bounty-operator]

command = "npx" args = ["-y", "https://bountyoperator.com/dl/bounty-operator-mcp.tgz"] env_vars = ["BOUNTY_OPERATOR_TOKEN", "OPENROUTER_API_KEY"] tool_timeout_sec = 1200

A hosted review ends within 270 seconds. On a Bounty Operator model it runs for up to 15 minutes: the first request after a quiet spell can wait up to 10 minutes for the model to start, and the progress message says so while it does. Codex stops a tool call after 60 seconds by default, so keep the tool_timeout_sec line.

The token and the key are read from the environment. Neither is a tool argument, and both are removed from every result.

The remote endpoint takes the same two values as headers: Authorization: Bearer and X-Provider-Key: . The commands for each client are at bountyoperator.com/mcp.

For Novita AI, set NOVITA_API_KEY locally (and include it in Codex's env_vars) and call run_review with "provider": "novita". Create a key at Novita Key Management. The default is deepseek/deepseek-v4.1-flash; zai-org/glm-5.3-flash is also listed. Both local and remote list_profiles expose the provider and its models. On the remote endpoint, pass the Novita key in X-Provider-Key, not as a tool argument. Novita bills model usage separately; selecting it does not change the app's default provider or enable a fallback. Signup attribution is never sent with inference requests.

For Nebius Token Factory, set NEBIUS_API_KEY locally (and include it in Codex's env_vars) and call run_review with "provider": "nebius". Create a key at Nebius Token Factory. The default is deepseek-ai/DeepSeek-V4.1-Flash; zai-org/GLM-5.3-Flash and moonshotai/Kimi-K3 are also listed. Requests ask Nebius not to store the answer. On the remote endpoint, pass the Nebius key in X-Provider-Key, not as a tool argument. Nebius bills model usage separately; selecting it does not change the app's default provider or enable a fallback.

Environment

What the server reaches

• list_profiles, prepare_review, run_gauntlet_plan and build_packet run on your machine with no network call.
• account and run_review call https://bountyoperator.com with your token. run_review sends the files and your provider key. The service adds the method of the profile, passes the request and the key to the provider for that one request and returns the review. It stores no files, keys or review text.
• The server reads the files you name under the working directory. It writes nothing to disk and runs no code from the files.

When a call fails

A failed call returns isError with one JSON object: