AgentPing

Get started

AgentPing protects each computer where people use AI. Getting a team going takes three steps.

  1. Create your organisation. Ask us for a workspace (the first 5 seats are free). Its first admin gets an email invitation and signs in to the console.
  2. Invite people. In the console, open Settings › People and invite each person as a Device user (their computer only), Analyst (read-only console) or Admin.
  3. Install on each computer and sign in with the organisation's name. Protection starts at once; findings appear in the console within minutes.

Until a computer signs in, AgentPing still protects it locally. Nothing is shared with anyone.

Install

The installer adds the AgentPing engine and menu bar app, then connects every supported AI tool it finds on the computer.

$ curl -fsSL https://github.com/madsquarebs/homebrew-agentping/releases/latest/download/install.sh | sh

The installer runs agentping setup, which turns on everything below: hooks for each coding agent found, the LLM proxy, the browser extension's connection and MCP protection. Run it again any time; agentping setup --remove undoes it.

Rolling out to many computers? Use your device management (Jamf, Intune, Kandji, Group Policy) to run the installer. agentping browser policy writes the browser policy that installs the extension on managed computers.

Sign in a computer

In the AgentPing menu bar app, enter your organisation's name and press Continue. The browser opens your organisation's sign-in (password, Google, Microsoft or your company's single sign-on). Check the code matches the one in the app and press Connect this computer.

From a terminal:

$ agentping login --server https://app.agentping.dev --org "Your company"
Opening your browser to sign in to AgentPing.
Sign in to Your company. Check the page shows code BCDF-GHJK, then press Connect.
Connected to Your company.

The computer joins as yours and takes your seat. Your organisation's policy applies from the next sync. agentping logout or Sign out in the app disconnects it; protection carries on locally.

Coding agents

AgentPing connects to each coding agent through its official hooks, so it sees prompts, files the agent opens, shell commands and tool results before they're sent.

AgentWhereWhat's checked
Claude CodeTerminal, VS Code, JetBrainsPrompts, file reads, searches, shell commands, edits, MCP calls, tool output
OpenAI CodexCLIPrompts, shell commands, tool output
CursorEditorPrompts and attachments, file reads, shell commands, MCP calls
GitHub CopilotVS Code agent mode, Copilot CLIPrompts with attached context, tool calls, tool output
$ agentping hooks install --agent auto     # every agent found (setup does this)
$ agentping hooks status                  # what's connected

When an agent tries to read a sensitive file such as .env, cloud credentials or an SSH key, the read is stopped and the agent is told why. If it runs a command that prints a file (cat, head, grep, jq…), the file is checked first, the same way.

LLM proxy

Hooks can stop a send but can't always rewrite it. The local proxy can: agents' requests to the model go through it, and sensitive values are replaced with placeholders like «REDACTED:aws_access_key_id:3f2a» so the agent keeps working. It also measures token usage and cost for the console.

$ agentping proxy service install     # starts at login, for every app (setup does this)
$ agentping proxy env                 # the variables, for a single shell
export ANTHROPIC_BASE_URL=http://127.0.0.1:8787/anthropic
export OPENAI_BASE_URL=http://127.0.0.1:8787/openai/v1

Restart editors and terminals after installing so they pick up the proxy. Your company already has an LLM gateway? Point the proxy at it with --upstream anthropic=https://gateway.example.com.

Browser

The AgentPing extension checks what people type, paste and upload in AI chats before the page sends it. Sensitive values are replaced in the message box, uploads are cleaned (values blacked out) or stopped, and the person sees a short banner saying what happened.

  • Browsers: Chrome, Edge, Brave, Arc, Opera, Vivaldi and Firefox.
  • Sites: ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, DeepSeek and Grok.
  • Uploads: text, CSV, PDF, Office documents, archives and images (screenshots are read with on-device text recognition).
$ agentping browser install     # connect the extension to the engine (setup does this)
$ agentping browser status

The extension never sends text anywhere but to the AgentPing engine on the same computer.

MCP servers

AgentPing wraps the MCP servers your agents use, local and remote, so tool calls carrying secrets or personal data are cleaned before they reach the server.

$ agentping mcp install --include-remote
$ agentping mcp status

Files and screenshots

Files are inspected wherever they're sent: uploads in the browser, files a coding agent opens, attachments. AgentPing extracts the text of PDFs, Word, Excel and PowerPoint documents and archives, and reads images and scanned PDFs with the operating system's own text recognition (macOS Vision, Windows OCR), on the device.

When a file holds something sensitive, AgentPing sends a cleaned copy with the values blacked out, or stops the upload if your policy says so. Files it can't read (encrypted, too large) can be blocked or let through with a warning.

What's detected

KindExamplesDefault
SecretsCloud keys (AWS, Azure, Google Cloud), GitHub, GitLab, Slack, Stripe, OpenAI and Anthropic keys; private keys; passwords in text; database connection strings; 140+ formatsReplaced
Payment cardsCard numbers (checksum-validated)Replaced
Identity numbersUS SSN, Aadhaar, PAN, passports, driving licences, national IDs and tax numbers from 40+ countries (checksum-validated)Replaced
Personal dataEmail addresses, phone numbers, IBANsWarned
Sensitive files.env, cloud credentials, SSH keys, production configs, keystoresRead stopped
Agent mistakesPackages that don't exist, versions that were never published, unbacked claimsBlocked

agentping rules lists every detector on a computer.

Policies

A policy says what happens to each kind of finding: allow, audit, warn, redact (replace) or block. The defaults replace secrets, card numbers and ID numbers, warn about other personal data, and block agent mistakes.

Organisation policy

Admins edit the policy in the console under Policies: a form for the common choices and YAML for the rest. Simulate replays recent findings under the draft before anyone is affected. Published policies reach computers at their next sync and keep working offline.

mode: enforce              # monitor: record only · warn: tell, don't stop · enforce
standard:
  sensitive_data: redact   # or block: stop the send instead of replacing
policy:
  rules:
    - match: { category: pii, surface: prompt.browser }
      action: block        # no personal data in browser chats

Per repository

A repository can be stricter with .agentping.yml at its root, in the same format. The organisation's policy is a floor: each finding gets the stricter of the two actions.

New organisations often start in monitor mode for a week to see what would be stopped, then switch to enforce.

Console

PageWhat it's for
OverviewFindings by severity, what was prevented, coverage, agents in use, spend
FindingsEvery finding with its risk and outcome; filter by severity, status, project or device; triage
Projects and devicesFindings, AI requests and cost per project (browser chats are “Browser”) and per device, with whose computer it is
CoverageWhich computers are protected, which agents are connected, which run an old policy
Agent qualityAgent mistakes and how often agents correct themselves
PoliciesEdit, simulate and publish the organisation policy
Usage and costRequests, tokens and cost by model and day
Audit logEvery change by admins and AgentPing staff
SettingsPeople, single sign-on, alerts to Slack and API keys

People and sign-in

People sign in with their work account: password, Google, Microsoft, passwordless codes or your company's single sign-on, with multi-factor sign-in and password reset built in.

RoleConsoleComputer
AdminEverything, including policies, devices and peopleIf they sign one in
AnalystRead only: findings, devices and reportsIf they sign one in
Device userNo consoleYes (takes a seat)

Single sign-on and directory sync: in Settings › Single sign-on and directory, your IT team connects Okta, Microsoft Entra ID, Google Workspace or any SAML/OIDC provider, and your user directory (SCIM) so joiners and leavers are added and removed automatically. Each opens a guided setup.

API keys (Settings) are for scripts and integrations, such as pulling findings into a SIEM. They can't sign in to the console.

Scanning in CI

The same detectors can check a repository before code is merged:

$ agentping scan --fail-on restricted .
config/settings.py:12  AWS access key ID (restricted)
1 finding · exit code 1

--json gives machine-readable output; --fail-on confidential is stricter, none only reports.

Command line

CommandWhat it does
agentping setupTurn on every integration found (the installer runs it)
agentping login --server URL --org NAMESign this computer in to your organisation
agentping logoutSign out; local protection carries on
agentping statusProtection, connection and policy on this computer
agentping dashboardThis computer's own dashboard (a local page, no sign-in)
agentping activity -fFollow what's being checked, live
agentping findingsRecent findings on this computer
agentping hooks install|statusCoding agent hooks
agentping proxy service installThe LLM proxy, at login
agentping browser install|statusThe browser extension's connection
agentping mcp install|statusMCP server protection
agentping scan PATHCheck files and folders
agentping usage --days 30Tokens and cost on this computer
agentping doctorFind and explain problems
agentping updateInstall the latest signed release

Self-hosting

Enterprise self-hosted runs the console and API in your own cloud. It's one container plus Postgres 16 or later; database migrations run when it starts. Contact sales for a licence.

$ AGENTPING_TOKEN_KEY=$(openssl rand -hex 32) docker compose up -d
SettingPurpose
DATABASE_URLPostgres connection. Each workspace's rows are isolated with row-level security.
AGENTPING_TOKEN_KEYRandom secret that signs device tokens
AGENTPING_PUBLIC_URLThe service's https address, as computers and browsers reach it
AGENTPING_RETENTION_DAYSHow long events and findings are kept (default 365)
WORKOS_CLIENT_ID, WORKOS_API_KEYSign-in, single sign-on and directory sync

Run the image's maintain command daily: it prepares storage, applies retention and clears expired sessions. Computers then sign in with --server set to your address.

Troubleshooting

  • Something isn't protected: run agentping doctor; it lists what's missing and how to fix it.
  • Claude Code blocks instead of replacing: it isn't using the proxy yet. Restart the editor or terminal after setup.
  • A finding is wrong: mark it False positive in the console; detectors are tuned from these.
  • “No organisation called …”: check the name with your admin; it's shown in the console under Settings.
  • Still stuck? Email support@agentping.dev.