Crabeye Mcp Bridge — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Crabeye Mcp Bridge (Plugin) and scored it 15/100 (red). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 2 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
This plugin references the AWS credentials file or the access-key fields stored inside it (* config could otherwise interpolate any env var…). Those are long-lived keys with broad cloud access, so any code that reads them can hand your whole AWS account to whatever it contacts next.
creds = open(os.path.expanduser("~/.aws/credentials")).read()
requests.post(url, data={"creds": creds})# let the SDK resolve credentials; never read or transmit the file yourself
import boto3
s3 = boto3.client("s3")This plugin references the AWS credentials file or the access-key fields stored inside it (for (const name of ["AWS_SECRET_ACCESS_KEY", "PA…). Those are long-lived keys with broad cloud access, so any code that reads them can hand your whole AWS account to whatever it contacts next.
creds = open(os.path.expanduser("~/.aws/credentials")).read()
requests.post(url, data={"creds": creds})# let the SDK resolve credentials; never read or transmit the file yourself
import boto3
s3 = boto3.client("s3")Every scanned point with the score it earned and what moved between them.
Score fell 85 points between these scans.
The primary manifest — the file an agent reads to learn what this artifact does.
One MCP connection for all your tools — with discovery, namespacing, and execution policies.
Every MCP server you add to your AI assistant means another connection, another set of tool definitions injected into the context window, and no way to search or control them centrally. Wire up ten servers with a hundred tools each and your assistant is burning tokens on a thousand tool schemas before the conversation even starts — most of which it will never call.
crabeye-mcp-bridge consolidates all your upstream MCP servers behind a single STDIO interface and exposes exactly two tools to the assistant: search_tools and run_tool. Tools from every server are discovered, namespaced, and indexed at startup, but none of them touch the context window until the assistant actually searches for them. You can have a thousand tools ready to go without bloating the context, with fuzzy search to find them and per-tool execution policies to control what runs freely, what needs approval, and what is blocked.
crabeye-mcp-bridge topology: one STDIO connection from an AI assistant to the bridge exposes two tools — search_tools and run_tool — while the bridge fans out to upstream MCP servers like Linear, GitHub, Slack, Sentry, and Figma. Tool schemas stay out of the context window until a search hits.
The fastest way to get started is with init, which discovers your MCP client configs and sets up the bridge automatically:
npx @crabeye-ai/crabeye-mcp-bridge initThis scans for config files from Claude Desktop, Cursor, VS Code Copilot, Windsurf, Zed, Cline, Roo Code, opencode, and Continue.dev, lets you pick which ones to use, and optionally injects the bridge entry. After that, just run npx @crabeye-ai/crabeye-mcp-bridge — no --config flag needed.
opencode and Continue.dev are surfaced as detect-only: their config schemas don't fit the rename-and-inject pipeline, so init prints a manual snippet for them instead of rewriting the file.
To undo, run npx @crabeye-ai/crabeye-mcp-bridge restore.
If you prefer to set things up manually, say your MCP client config looks like this today:
{
"mcpServers": {
"linear": {
"command": "npx",
"args": ["-y", "@anthropic/linear-mcp-server"]
},
"github": {
"command": "npx",
"args": ["-y", "@anthropic/github-mcp-server"],
"env": {
"GITHUB_TOKEN": "ghp_..."
}
}
}
}First, store your secrets in the encrypted credential store:
crabeye-mcp-bridge credential set github-pat ghp_abc123Then rename mcpServers to upstreamMcpServers, add the bridge, and replace hardcoded tokens with ${credential:key} references:
{
"mcpServers": {
"bridge": {
"command": "npx",
"args": ["-y", "@crabeye-ai/crabeye-mcp-bridge", "--config", "/path/to/this/file.json"]
}
},
"upstreamMcpServers": {
"linear": {
"command": "npx",
"args": ["-y", "@anthropic/linear-mcp-server"]
},
"github": {
"command": "npx",
"args": ["-y", "@anthropic/github-mcp-server"],
"env": {
"GITHUB_TOKEN": "${credential:github-pat}"
}
}
}
}That's it. Your AI assistant now has access to all tools from all configured servers through a single connection. The bridge automatically excludes itself from mcpServers to avoid recursion, so pointing --config at the same file is safe.
The bridge also reads upstreamServers (shorthand), servers (VS Code Copilot), and context_servers (Zed) as input keys. See docs/configuration.md for the full priority order and self-exclusion rules.
Alternatively, you can add the bridge alongside your existing mcpServers entries without renaming anything — the bridge will pick up the other servers from mcpServers automatically (excluding itself). Disable the other MCP servers in your client so the assistant uses the bridge as the single entry point.
Register the bridge's repo as a Claude Code marketplace, then install the plugin:
/plugin marketplace add crabeye-ai/crabeye-mcp-bridge
/plugin install crabeye-mcp-bridge@crabeyeClaude Code auto-registers the bridge as an MCP server. If you haven't configured upstreams yet, the bridge's startup error will surface in Claude Code's MCP server log — run npx @crabeye-ai/crabeye-mcp-bridge init in a terminal to discover and import existing MCP server configs from your other clients.
search_tools, run_tool) instead of N×M tool definitions in context. See docs/how-it-works.md.${credential:key} templates, and a one-shot auth <server> OAuth flow for HTTP upstreams. See docs/auth.md.always / prompt / never), rate limiting, discovery modes. See docs/policies.md.init, restore, credential, daemon, --validate. See docs/cli.md.MIT
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.