Skip to main content
Glama
AgentsID Guard

AgentsID Guard

Official
by AgentsID-dev

Your AI agent has access to your shell, file system, database, git repos, and the internet. Right now, nothing controls what it can do with any of them.

AgentsID Guard fixes that. 11 tools across 5 categories — shell, files, database, git, HTTP — every operation validated against per-agent permission rules before execution.

How It Works

Agent: shell_run("ls -la /src")
  → Classified: shell.read.ls
  → AgentsID: ALLOWED ✓
  → Executes normally

Agent: shell_run("rm -rf /data")
  → Classified: shell.danger.rm
  → AgentsID: BLOCKED ✗
  → Never executes

Agent: db_query("DROP TABLE users")
  → Classified: db.danger.ddl
  → AgentsID: BLOCKED ✗
  → Never executes

Agent: git_run("push origin main")
  → Classified: git.write.push
  → AgentsID: BLOCKED ✗
  → Never executes

Related MCP server: gov-mcp

Quick Start

1. Install

npx @agentsid/guard

2. Get your keys

Sign up at agentsid.dev/dashboard — free tier: 25 agents, 10K events/month.

3. Add to Claude Code

claude mcp add guard \
  -e AGENTSID_PROJECT_KEY=aid_proj_your_key \
  -e AGENTSID_AGENT_TOKEN=aid_tok_your_token \
  -- npx @agentsid/guard

4. Set permissions

npx agentsid register-agent --name "my-agent" \
  --permissions "shell.read.*" "file.read" "file.list" "git.read.*" "http.get"

Tools

AgentsID Guard exposes 11 MCP tools:

Tool

What it does

Permission pattern

shell_run

Execute a shell command

shell.read.*, shell.write.*, shell.danger.*, shell.admin.*

file_read

Read a file

file.read

file_write

Write/create a file

file.write

file_delete

Delete a file

file.delete

file_list

List directory contents

file.list

file_info

Get file metadata

file.info

db_query

Run a SQL query

db.read, db.write.*, db.danger.*

git_run

Run a git command

git.read.*, git.write.*, git.danger.*

http_request

Make an HTTP request

http.get, http.post, http.put, http.delete

check_permission

Check if an action would be allowed

—

list_categories

List all permission categories

—

Permission Categories

Shell

Pattern

Commands

Risk

shell.read.*

ls, cat, grep, find, ps, df, curl, ping

Safe

shell.write.*

mkdir, touch, cp, mv

Moderate

shell.danger.*

rm, chmod, chown, kill

High

shell.admin.*

sudo, docker, apt, npm, systemctl

Critical

Files

Pattern

Operations

Risk

file.read

Read file contents

Safe

file.list

List directories

Safe

file.info

File metadata

Safe

file.write

Create/write files

Moderate

file.delete

Delete files

High

Database

Pattern

Statements

Risk

db.read

SELECT, SHOW, DESCRIBE, EXPLAIN

Safe

db.write.insert

INSERT

Moderate

db.write.update

UPDATE

Moderate

db.write.create

CREATE

Moderate

db.danger.delete

DELETE

High

db.danger.ddl

DROP, TRUNCATE, ALTER

Critical

Git

Pattern

Commands

Risk

git.read.*

status, log, diff, branch, show, blame

Safe

git.write.*

add, commit, push, pull, merge, checkout, stash

Moderate

git.danger.*

reset, force-push

Critical

HTTP

Pattern

Methods

Risk

http.get

GET

Safe

http.post

POST

Moderate

http.put

PUT

Moderate

http.delete

DELETE

High

Permission Examples

Read-only research agent:

shell.read.*    → allow
file.read       → allow
file.list       → allow
db.read         → allow
git.read.*      → allow
http.get        → allow

Developer agent (read + write, no destructive):

shell.read.*    → allow
shell.write.*   → allow
file.read       → allow
file.write      → allow
file.list       → allow
db.read         → allow
db.write.*      → allow
git.read.*      → allow
git.write.*     → allow
http.get        → allow
http.post       → allow

Full access with approval gates:

shell.read.*    → allow
shell.write.*   → allow
shell.danger.*  → allow (requires approval)
shell.admin.*   → allow (requires approval)
file.*          → allow
db.read         → allow
db.write.*      → allow
db.danger.*     → allow (requires approval)
git.*           → allow
http.*          → allow

Environment Variables

Variable

Required

Description

AGENTSID_PROJECT_KEY

Yes

Your AgentsID project key

AGENTSID_AGENT_TOKEN

Yes

Agent token for permission checks

AGENTSID_BASE_URL

No

API URL (default: https://agentsid.dev)

GUARD_CWD

No

Working directory (default: cwd)

GUARD_TIMEOUT

No

Command timeout in ms (default: 30000)

GUARD_ALLOWED_DIRS

No

Comma-separated allowed directories for file ops (default: cwd)

GUARD_DB_URL

No

Database connection string for db_query (postgresql://, mysql://, or .db path)

Security

  • Deny-first — unknown commands and tools are blocked by default

  • Path containment — file operations restricted to allowed directories

  • Fail-closed — network errors to AgentsID result in denial, not bypass

  • Audit trail — every allow and deny logged to AgentsID's tamper-evident hash chain

  • No shell injection — commands executed via execSync with no shell interpolation of user input in tool arguments

Dashboard

Every operation appears in your AgentsID dashboard:

  • Which agent ran which command

  • Whether it was allowed or denied and why

  • Full audit trail across all 5 categories

License

MIT

Available Tools

50 tools
aws_runA

Run an AWS CLI command. Read ops (ls, describe) require aws.read.{service}. Write ops require aws.write.{service}. Destructive ops (delete, terminate) require aws.danger.{service}.

Inputs: command (required).

check_permissionA

Check if a tool/action would be allowed without executing it.

Inputs: tool (required).

container_inspectA

Get detailed info about a Docker container. Requires container.read permission.

Inputs: id (required).

container_listA

List Docker containers. Requires container.read permission.

Inputs: all.

container_logsB

View Docker container logs. Requires container.read permission.

Inputs: id (required), tail.

container_removeA

Remove a Docker container. Requires container.danger permission.

Inputs: id (required), force.

container_startA

Start a stopped Docker container. Requires container.write permission.

Inputs: id (required).

container_stopA

Stop a running Docker container. Requires container.write permission.

Inputs: id (required).

cron_addA

Add a cron job. Requires cron.write permission.

Inputs: command (required), schedule (required).

cron_listA

List current user's cron jobs. Requires cron.read permission.

Inputs: no named parameters.

cron_removeA

Remove cron jobs matching a pattern. Requires cron.danger permission.

Inputs: pattern (required).

db_queryA

Run a SQL query. SELECT/SHOW requires db.read. INSERT/UPDATE requires db.write.. DELETE/DROP requires db.danger.. Supports PostgreSQL, MySQL, SQLite.

Inputs: query (required), database.

disk_usageB

Get disk usage for a path. Requires system.read.

Inputs: path.

env_getA

Get a specific environment variable. Secret-named vars (KEY, TOKEN, PASSWORD) require env.read.secret permission.

Inputs: name (required).

env_listA

List environment variables (secrets auto-filtered). Requires env.read permission.

Inputs: no named parameters.

env_setA

Set an environment variable. Requires env.write permission.

Inputs: name (required), value (required).

file_deleteA

Delete a file. Requires file.delete permission.

Inputs: path (required).

file_infoA

Get file metadata (size, type, permissions, timestamps). Requires file.info permission.

Inputs: path (required).

file_listA

List files in a directory with size and type info. Requires file.list permission.

Inputs: path.

file_readA

Read a file's contents. Restricted to allowed directories. Requires file.read permission.

Inputs: path (required).

file_writeA

Write content to a file. Creates parent directories if needed. Requires file.write permission.

Inputs: path (required), content (required).

git_runA

Run a git command. Read operations (status, log, diff) require git.read.. Write operations (commit, push) require git.write..

Inputs: command (required).

http_requestB

Make an HTTP request. GET requires http.get. POST requires http.post. PUT requires http.put. DELETE requires http.delete.

Inputs: url (required), body, method, headers.

kubectl_runA

Run a kubectl command. Read ops (get, describe, logs) require k8s.read.. Write ops (apply, create) require k8s.write.. Destructive ops (delete, drain) require k8s.danger.*.

Inputs: command (required).

list_categoriesA

List all permission categories and what they control.

Inputs: no named parameters.

log_readB

Read last N lines of a log file. Requires log.read.

Inputs: path (required), lines.

memory_usageB

Get memory usage breakdown. Requires system.read.

Inputs: no named parameters.

network_dnsA

DNS lookup for a domain. Requires network.read permission.

Inputs: domain (required), type.

network_pingA

Ping a host. Requires network.read permission.

Inputs: host (required), count.

network_portsA

List listening network ports. Requires network.read permission.

Inputs: no named parameters.

network_tracerouteA

Traceroute to a host. Requires network.read permission.

Inputs: host (required).

package_infoC

Get info about a package. Requires package.read.

Inputs: name (required), manager.

package_installB

Install a package. Requires package.danger.install (high risk).

Inputs: name (required), manager.

package_listA

List installed packages. Requires package.read.

Inputs: manager.

process_infoA

Get detailed info about a specific process. Requires process.read permission.

Inputs: pid (required).

process_killA

Send a signal to a process. Requires process.kill permission.

Inputs: pid (required), signal.

process_listA

List running processes sorted by CPU usage. Requires process.read permission.

Inputs: no named parameters.

secrets_anomaliesA

Detect vault anomalies: expired secrets, broad scopes, stale keys, unusual access patterns. Requires secrets.read.

Inputs: no named parameters.

secrets_auditA

View access log for secrets — who accessed what, when, for what purpose. Requires secrets.read.

Inputs: name.

secrets_deleteB

Delete a secret from the vault. Requires secrets.danger.

Inputs: name (required).

secrets_expireA

Set a time-to-live on a secret. After expiry, proxy injection will refuse to use it. Requires secrets.write.

Inputs: name (required), ttlMinutes (required).

secrets_historyB

View rotation history for a secret. Requires secrets.read.

Inputs: name (required).

secrets_injectA

Make an HTTP request with vault secrets injected server-side. The agent NEVER sees the raw secret — Guard substitutes placeholders like {{API_KEY}} with the real value, makes the call, and scrubs secrets from the response. Requires secrets.proxy.

Inputs: url (required), secretMappings (required), body, method, +1 more.

secrets_listA

List secret names (values never exposed). Requires secrets.read.

Inputs: no named parameters.

secrets_scopeA

Restrict which tools can use a secret. E.g., scope an API key to only http.post calls. Requires secrets.write.

Inputs: name (required), allowedTools (required).

secrets_storeA

Store a secret in the encrypted vault. Per-secret derived encryption keys. Requires secrets.write.

Inputs: name (required), value (required), metadata.

shell_runA

Run a shell command. Validated against AgentsID permissions. Read commands (ls, cat, grep) typically allowed. Destructive commands (rm, chmod) blocked unless permitted.

Inputs: command (required).

ssh_runA

Run a command on a remote host via SSH. Permission scoped per host (ssh.hostname). Requires SSH keys configured.

Inputs: host (required), command (required), port, user.

system_infoA

Get system overview: OS, CPU, memory, disk, uptime. Requires system.read.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 50 tools

    Disambiguation4/5

    Most tools are clearly separated by domain and action, but several run commands (shell, git, aws, kubectl, ssh) could be confused. Descriptions help by scoping each to a specific system, but the overlapping purpose makes misselection possible.

    Naming Consistency5/5

    All tool names are lowercase with underscores and consistently use a domain prefix plus a noun or verb (e.g., file_read, env_list, secrets_delete). The pattern is highly predictable and uniform across the set.

    Tool Count2/5

    With 50 tools, the server is heavy and covers many distinct subsystems. While each tool has a purpose, the count exceeds typical MCP server scopes and may overwhelm agents with too many options.

    Completeness4/5

    The tool set provides broad lifecycle coverage for secrets, files, processes, containers, and more. Minor gaps exist (e.g., no explicit update for cron jobs), but agents can work around them without dead ends.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      A Model Context Protocol (MCP) server that enables secure execution of shell commands with a dynamic approval system, audit logging, and command revocation.
      42
      Apache 2.0
    • A
      license
      Not graded
      quality
      D
      maintenance
      An MCP server that enforces runtime governance on AI agent actions — file access, command execution, delegation chains, and permission escalation.
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      A least-privilege enforcement proxy for MCP servers. It sits between MCP clients and upstream servers, enforcing tool policies, hiding denied tools, requiring human approval for risky actions, and providing a structured audit trail.
      MIT

    Appeared in Searches