Agent Canary 3.0 is an open-source MCP and SDK security toolkit for AI agents. It provides decoy tools, canary tokens, a guarded execution gateway, local self-tests, and an optional risk-policy console.

Current mainline: 3.0.0 · MIT · Node.js 20+

• GitHub release
• Project website and test results
• Getting started from a pinned source snapshot
• Chinese documentation
• Architecture and acceptance report
• Source and issues

The npm package is published to GitHub Packages. Configure the @dorianchn scope, then authenticate with a GitHub personal access token (classic) that has read:packages:

npm config set @dorianchn:registry https://npm.pkg.github.com
npm login --scope=@dorianchn --auth-type=legacy --registry=https://npm.pkg.github.com
npm install --global @dorianchn/[email protected]
agent-canary --help

If authentication fails, check the status before changing credentials:

• 401: confirm the token is a personal access token (classic) with read:packages; enter your GitHub username and token when npm login prompts. npm 9 and newer require --auth-type=legacy for this prompt-based login.
• 403: confirm your account can access the package and authorize the token for SSO if the organization requires it.
• 404 or a missing version: check the package name and version, then run npm config get @dorianchn:registry and confirm it prints https://npm.pkg.github.com.

GitHub Packages requires authentication for public packages too. Never commit a token to the repository or paste it into logs or chat; see GitHub's npm registry authentication guide. To build from a pinned source snapshot without logging in to the package registry, follow Getting started.

The Docker image is published to GitHub Container Registry:

docker pull ghcr.io/dorianchn/shanchuanzhi-agent-canary:3.0.0
docker run --rm ghcr.io/dorianchn/shanchuanzhi-agent-canary:3.0.0 --help

To run the MCP server over stdio, use agent-canary serve --stdio or run the container with -i and the same arguments. A minimal MCP client entry after global installation is:

{
  "mcpServers": {
    "agent-canary": {
      "command": "agent-canary",
      "args": ["serve", "--stdio"]
    }
  }
}

The client-specific config file location varies. Agent Canary protects only callbacks routed through this server; it does not intercept the client's native tools.

import { createTrustedExecutionGateway } from "@dorianchn/agent-canary/sdk";

For a runnable example that exercises an allowed read and a denied write, see the SDK example in the Chinese README. The repository's automated test executes that example against the local Gateway.

The host must register the callbacks it wants the gateway to protect. Agent Canary does not automatically control unrelated tools, processes, or native host capabilities.

The R45 integration adds a Cline SDK Plugin and a fail-closed adapter for the upstream Filesystem MCP server. In the tested Cline Agent Runtime path, Cline registers only the protected Agent Canary tools; its beforeTool hook skips other native or direct MCP tools before their executors run. File content reads use the existing Gateway-held file capability. Directory listing and metadata requests are checked by the Gateway before the adapter calls the upstream server in read-only mode. Writes are denied.

The recorded end-to-end run passed one real stdio Filesystem MCP proxy test and one Cline Agent Runtime test. The run used synthetic files and a deterministic in-process tool-calling model; it did not call an external model API. Raw events, Gateway decisions, pinned upstream revisions, coverage limits, and setup instructions are in the R45 integration report and integrations/.

This verification covers the pinned Cline SDK Plugin and Agent Runtime route only. The Cline CLI executable and IDE applications were not run. The integration is application-level routing, not OS confinement; do not expose direct Filesystem MCP or native tools outside the plugin boundary and assume they are protected. E4 and the broader model matrix were not run.

The published acceptance record for the frozen 3.0 source reports:

• Community automated suite: 100/100 passed.
• Windows 11 same-user installation flow: 9 steps passed; the installer is unsigned. Clean Windows VM validation remains BLOCKED.
• Ubuntu portable-package checks: SDK types, doctor, effects, uninstall, and repeat uninstall passed in the recorded WSL2 environment.
• Core and wrapper CI: 9/9 and 2/2 jobs passed for the recorded commits.
• Real local Qwen tool-selection run: 48 trials. Direct forbidden effects were 12/12 in the unprotected group and 0/12 in the guarded group; guarded calls actively refused 12/12. README-indirect tasks produced 0/4 effects in both groups, so they do not show an active prevention gain. Normal tasks completed 8/8 in both groups.
• Independent reproduction E4: BLOCKED. No independent operator and environment have reproduced the frozen build. Same-host tests and log review do not count as E4.

These results apply to the listed source, tests, and environment. They do not establish global protection, zero false positives, or complete TOCTOU elimination. A non-cooperative callback may still cause a late side effect. See security limitations and the full acceptance report.

npm ci
npm run verify

The verification command builds the TypeScript sources, runs the automated suite, performs offline acceptance and SDK type checks, verifies the source manifest, and smoke-tests the packed npm package.

The demonstration video will be added after it is uploaded.