Security for AI agents, built in the open.

Coding agents now run commands, install packages and read whatever they are pointed at. HOL Guard reviews every action before it runs, the Plugin Registry scores plugins before you install one, and CIGAR gives agents cited context within a budget. All of it open source and free to run.

  • Apache-2.0, source on GitHub
  • Runs locally, works offline
  • 13 coding agents supported
  • Open HCS standards

HOL Guard · review

Claude Code

$ curl -fsSL get.devtool.sh | sh

Runs a script from the internet in your shell.

BlockAllow onceAlways allow

Plugin Registry

Every listing scanned

CIGAR · context

local

compile("how does token refresh work?")

Budget3,412 / 4,000 tokens

verified7 sources · 4 cited

Guard downloads
2.5M+
PyPI, Guard and plugin-scanner
GitHub stars
6.2K+
Across HOL repositories
Latest Guard release
v3.58.0
Released Oct 27, 2026
Guard works with
Claude CodeCodexCursorGemini CLIOpenCodeand more

Sponsor HOL

Let's secure our future.

AI agents will write and run more of the world's software every year. The security underneath them should be open and owned by no single company. Sponsorship keeps it that way.

What a sponsorship fundsfrom $5,000 a year
  1. 01

    Safer command execution

  2. 02

    Plugin and MCP supply-chain security

  3. 03

    Reproducible testing and compatibility

  4. 04

    Open implementation guidance

The plan

Five places the agent era needs open infrastructure.

Agents act, install, read and find each other. HOL builds one open layer for each, on standards no single company owns.

  1. 01

    Where an agent acts.

    HOL Guard

    Reviews every command, package install and config change before it runs, on your machine.

    Explore HOL Guard
  2. 02

    Before an agent installs.

    HOL Plugin Registry

    Shows who built a plugin, MCP server or skill, whether the owner is verified, and a trust score you can compare.

    Explore HOL Plugin Registry
  3. 03

    What an agent reads.

    CIGAR

    Turns your documents and code into cited context that fits a token budget, with authority that stays in your app.

    Explore CIGAR
  4. 04

    How agents find each other.

    Registry Broker

    One API and one universal agent ID to search and resolve agents across registries and protocols.

    Explore Registry Broker
  5. 05

    The rules anyone can build on.

    HCS standards

    Open specifications for identity, messaging, registries and trust, written in public with partners.

    Explore HCS standards

01 · Where an agent acts

See the command before it runs.

HOL Guard sits between your coding agent and your machine. Routine work goes through. Anything new or risky stops and asks you first.

  1. $ git status && npm test

    Reads and runs your own tests

    Allowed
  2. $ npm install left-pad-utils

    New package, first install

    Asks you
  3. $ cat ~/.aws/credentials

    Reads cloud credentials

    Blocked
Install on macOS or Linux
$ set -o pipefail; curl -fsSL https://hol.org/guard/install.sh | bash -s -- --mode install --verify$ hol-guard init

Works with Codex, Claude Code, Copilot CLI and 10 more coding agents.

02 · Before an agent installs

Check a plugin before you install it.

The HOL Plugin Registry lists plugins, MCP servers and skills for agent tools. Each shows who built it, whether the owner is verified, and a trust score.

03 · What an agent reads

Give your agent the right context, with receipts.

CIGAR, the Context Intelligence Graph Agentic Runtime, turns documents and code into cited context that fits an agent's token budget. It runs locally through Python and TypeScript SDKs, with no model, account or API key.

  • Relevant by selection

    Deterministic ranking packs only what the question needs, from your own sources.

  • Bounded by design

    Your app sets the sources, scope and token budget. Relevance never overrides permission.

  • Traceable to the source

    Every context carries citations and a snapshot you can verify before an answer ships.

Latest, v0.14.1: independent agents share one local evidence graph while the host keeps authority, now with security fixes. Release notes

Example question

How should checkout retry a payment?

01 · Your sources

  • Retry policy

    Required evidence

  • Idempotency

    Required dependency

  • Incident report

    Supporting evidence

  • Old retry policy

    Declared counterevidence

  • Launch plan

    Outside this question

  • Refund guide

    Outside this question

02 · CIGAR

Context graph
  • Respects access
  • Follows relationships
  • Fits the budget
Rust core · local

03 · Cited context for your model

  1. Retry policy · required evidence[1]
  2. Idempotency · required dependency[2]
  3. Incident report · supporting evidence[3]
  4. Old retry policy · declared counterevidence[4]
Context ≤ 448 tokens64 reserved · 512 limit
An illustrative selection. Your application declares which sources are allowed, and every compiled context carries a snapshot you can verify.

04 · How agents find each other

Find any agent with one API.

The HOL Registry Broker searches and resolves AI agents across registries and protocols. Every result carries an HCS-14 universal agent ID and its trust signals.

Registry Broker API
GET /api/v1/search?q=translator&limit=2
200Example response, trimmed
{  "total": 12,  "hits": [    {      "name": "Polyglot",      "registry": "hashgraph-online",      "protocols": ["a2a", "mcp"],      "uaid": "uaid:aid:…"    },    …  ]}

Who builds it

Built in the open, with the companies that depend on it.

HOL is part of Hiero at the Linux Foundation Decentralized Trust. Everything here is developed in public with Partner Program members, in working groups on registries, payments, privacy and security, and inter-agent communication.

Founding partners
30+
Working groups formed
12+
GitHub stars across HOL repos
6.2K
Package downloads, last 30 days
329.3K

How the work is funded

Open source, with an open business model.

Nothing on this page sits behind a paywall. The work is paid for by the teams that run it, the companies that want to reach them, and sponsors. Every way to take part is listed here.

Fund the work · Sponsor

Sponsorship is how the open security work stays open.

Annual sponsorships pay for the engineering behind Guard, the registries, CIGAR and the standards, and let you scope the security problem you most want solved in public.

From$5,000a year

  • A linked company logo on an agreed HOL sponsor or partner page for 12 months
  • One jointly reviewed sponsor announcement on the HOL blog
  • Quarterly written updates on the open-source work your sponsorship supports
  • Strategic and initiative packages add quarterly conversations, introductions, and a scoped AI-security initiative

Three packages, each scoped in a written proposal before any payment.

Reach developers

Advertise

From $50 a month

Self-serve placements where developers choose agent tools.

  • Featured slots in the plugin and extension directories
  • Linked logos in HOL project READMEs, including HOL Guard
  • Exact monthly prices and self-serve checkout

Run it

HOL Guard

Free on your machine

Local protection and secret scanning never need an account. Guard Cloud adds what teams ask for.

  • Solo adds continuous GitHub leaked-secret monitoring and Cloud memory
  • Pro and Team add longer history, instant alerts and shared policy
  • Plugin owners can add an HOL Blue or Gold checkmark to their listing

Build the standards

Partner Program

By application

Co-author the specifications the agent ecosystem runs on, alongside the founding cohort of companies already building on them.

  • Four standing subcommittees and their working groups
  • Early access to reference implementations
  • A public partner profile on hol.org

FAQ

Questions people ask about HOL.

What is HOL?

HOL, short for Hashgraph Online, is open infrastructure for AI agents. It builds HOL Guard for agent security, the HOL Plugin Registry for finding plugins, MCP servers, and skills, CIGAR for cited, budgeted agent context, the Registry Broker for agent discovery, and the open HCS standards behind them.

What is HOL Guard?

HOL Guard is a local-first security layer for AI agents. On supported integrations it checks risky actions such as commands, tool calls, package installs, and configuration changes before they run, lets you allow, ask, or block them, and keeps a record of what happened. Install it with the one-line command on this page, then run hol-guard init.

Which AI tools does HOL Guard work with?

HOL Guard connects to Codex, Claude Code, Copilot CLI, Antigravity CLI, Cursor, OpenCode, Hermes, OpenClaw, Antigravity, Kimi, Grok, Pi, Z Code, or a custom agent. How much Guard can intercept depends on what each tool exposes, so check the Guard documentation at hol.org/guard/docs for the coverage of your tool.

What is the HOL Plugin Registry?

The HOL Plugin Registry is a public directory of plugins, MCP servers, and skills for AI agent tools. Each listing shows its developer, category, verification and trust signals, and scan results, so you can compare options before you install one.

What is CIGAR?

CIGAR, the Context Intelligence Graph Agentic Runtime, is open-source infrastructure for agent context. It turns documents and code into relevant, cited context that fits a token budget, keeps the allowed sources and authority with your application, and returns a verifiable snapshot with every compiled context. The Rust core runs locally through Python and TypeScript SDKs. See hol.org/cigar.

How do developers build on the HOL Registry Broker?

The Registry Broker is the HOL API for discovering and resolving AI agents across registries. Start with the documentation at hol.org/docs and the developer hub at hol.org/developers, then call the API described by the OpenAPI document at hol.org/openapi.json and the registry docs at hol.org/registry/docs.

What are the HCS standards?

The HCS standards are open specifications for agent identity, messaging, registries, and trust that HOL and its partners publish. Examples include HCS-10 for agent communication, HCS-14 for Universal Agent IDs, and HCS-26 for agent skills registries. The full list is at hol.org/standards.

How can my company support HOL's open-source work?

Sponsorship starts at $5,000 a year and includes a linked company logo on an agreed HOL sponsor or partner page for 12 months, a jointly reviewed announcement on the HOL blog, and quarterly written updates on the work it supports; see hol.org/sponsor. Advertising starts at $50 a month for a plugin directory placement and $50 a month for an extension directory placement; see hol.org/advertise. Organizations building on the standards can apply to join the HOL Partner Program and co-author them alongside the founding cohort.