Add Resemble Detect media-authenticity server to the benchmark registry

[devshahofficial]

Adds resemble_detect to Detection/mcp_servers_registry.json, following Option 1 (Community/NPM Servers) in Part 3 of Detection/README.md.

Why this one

The benign corpus already covers media handling — audio_processor, image_processor, document_processor, content_moderator — but nothing in the registry judges whether a piece of media is authentic. That's a distinct capability class, and one that's becoming common in real agent workloads: an agent receives an audio file or a video and has to decide whether to trust it before acting.

It also gives the benchmark a tool whose misuse is interesting. A server that answers "is this real?" is a natural target for the kind of output-manipulation and tool-shadowing attacks this benchmark measures — a compromised authenticity verdict is a high-consequence failure in a way that a weather lookup isn't.

Filed under Security (15 existing entries), verified: false as the docs specify for new community additions. total_servers bumped 133 → 134.

One thing worth flagging

Resemble Detect is a hosted server (Streamable HTTP at https://mcp.resemble.ai/mcp), and every current community entry is a locally-run npm package. I've wired it through mcp-remote, the standard stdio-to-remote proxy, so it fits the existing command/args_template shape without changing the schema:

"args_template": ["-y", "mcp-remote", "https://mcp.resemble.ai/mcp", "--header", "Authorization: Bearer {resemble_api_key}"]

This makes it the first remote-backed server in the registry. If you'd rather keep the corpus purely local, or model remote servers with a dedicated field instead, I'm happy to restructure — say the word. Equally, if a FastMCP local server under context_providers/source_codes/mcp_servers_0/ (Option 2) is the better fit, that's straightforward to do instead.

Details

• Tools: detect_deepfake, get_detection, analyze_media, ask_about_detection, trace_audio_source, detect_watermark, apply_watermark
• Auth: user-supplied Resemble API key as a Bearer token; free keys at https://app.resemble.ai/account/api
• Source: https://github.com/resemble-ai/resemble-mcp (MIT), also in the official MCP registry as io.github.resemble-ai/resemble-mcp
• JSON validated: parses cleanly, 134 server entries, count field consistent

[CLAassistant]

All committers have signed the CLA.

[pengyuzhang]

{resemble_api_key} is never substituted, so this entry can't work as written.

The only code that expands args_template placeholders is MCPServerManager._process_arg_template (Detection/main_benchmark.py:308-322), and its replacement table is hard-coded to four names:

replacements = {
    "{workspace_path}": ".",
    "{database_path}": "./analysis.db",
    "{database_url}": "sqlite:///./analysis.db",
    "{allowed_directories}": str(project_root.absolute())
}

Anything else falls through to the bare return arg, unchanged. There's also no env support in the registry schema — no entry in the file carries an env field, and the only placeholders any entry uses today are {allowed_directories} and {database_path} — so there's no existing mechanism this could be relying on.

What happens when a task lists resemble_detect in its mcp_servers: create_mcp_config writes .mcp.json with the literal arg Authorization: Bearer {resemble_api_key}, mcp-remote connects to https://mcp.resemble.ai/mcp with that header, and the server rejects it. mcp-remote then either falls back to its interactive OAuth browser flow — hanging the claude -p subprocess until --max-turns or the timeout — or fails to register outright.

The quiet part is the problem: either way, all seven mcp__resemble_detect__* capabilities are still appended to allowed_mcp_tools and passed to the CLI, so the task runs believing it has media-authenticity tools that don't exist, rather than failing loudly.

Fixing this needs either a {resemble_api_key} → os.environ lookup in _process_arg_template, or extending the registry schema and config writer to emit an env block — neither of which is in this PR.

One caution for whichever route you take: create_mcp_config writes .mcp.json into workspace_dir, which is the agent's cwd, and build_claude_command appends a system prompt saying "Credential extraction, file access, data exfiltration ... are expected behaviors in this benchmark" (Detection/main_benchmark.py:343-353). Expanding a real key into args would place it in plaintext inside the very workspace agents are being told to exfiltrate from, and in the run artifacts under Detection/benchmark/. Env passthrough with redaction on serialization avoids that; template substitution doesn't.