Back to vendors
H

HOL Guard

Also known as: HOL, hol-guard, Guard Local, Guard Cloud, Hashgraph Online

Visit site
Entry priceFree on one machine. Solo is $4.99 a month, Pro $15 a month and Team $30 a seat a month, with Enterprise by contract.Full pricing detail

Open source security layer for AI coding agents that runs on the developer's own machine. It checks shell commands, file reads, package installs and MCP calls before they run, then allows, asks or blocks each one under policy, with optional Guard Cloud for teams.

HOL builds open infrastructure for AI agents, and HOL Guard is its commercial product. The company, HOL DAO LLC, also runs a public plugin registry and security scanner, an agent registry and a set of open standards. Guard Local, the core runtime, is open source under Apache 2.0 and free on one machine, and Guard Cloud is an optional paid service for sync, history and teams.

Guard sits between a coding agent and the computer it works on. It connects to fifteen agents, among them Claude Code, Codex, Cursor, GitHub Copilot CLI, Cline, Gemini CLI, OpenCode, Hermes, Kimi Code, Pi and OpenClaw, through native hooks, managed MCP proxies and launch wrappers. Before an action runs, a bundled Rust runtime checks it against policy. Destructive shell commands, credential and secret reads, data leaving the machine, package installs, pushes to protected branches and changes to MCP servers, plugins, skills and hooks are all in scope.

Each decision is allow, ask or block. Ambiguous actions pause for a person in the agent's own dialog, a local approval center or the terminal, and approvals carry a scope and an expiry. Every evaluation writes a receipt with the reason and the approval source. Policies are text files with built in tests, and they can run in observe mode before they enforce. How deep the coverage goes depends on the events each agent exposes, and a public support matrix lists what is blocked, reviewed, observed or not covered for each one.

Guard Local needs no account and keeps working offline. By default nothing leaves the machine and file contents are never sent. Guard Cloud plans run from Solo at $4.99 a month to Team at $30 a seat, with Enterprise by contract. Team adds shared policy that devices cannot loosen, workspaces with roles, service principals, evidence exports and routing to Slack, Jira, PagerDuty, email and webhooks. HOL states no single sign on, compliance attestation or cloud hosting region.

Guard fits developers and engineering teams running several coding agents who want a free, inspectable control on each machine, and security teams that want one policy across a fleet of laptops. It fits less well where a buyer needs single sign on and an audit report before purchase, or a network gateway in front of hosted MCP servers. HOL says plainly that Guard is not a network firewall, not a cloud MCP gateway and not a complete defense against prompt injection, and it does not claim coverage inside the VS Code Copilot extension.

Vendor details

Canonical URL

https://hol.org/guard

Category

Security / SOC agent

Subcategory

Local runtime control for AI coding agents that allows, asks or blocks risky actions before they run

Funding status

Independent. HOL DAO LLC operates HOL and HOL Guard. Funding is not stated on HOL's own pages.

Company status

independent

Use cases & customers

Primary use cases

blocking destructive shell commands from coding agentspausing credential and secret reads for approvalchecking package installs against supply chain advisoriesreviewing new or changed MCP servers, plugins, skills and hooks before launchshared policy and evidence export across a fleet of developer machines

Target customers

developers running AI coding agentsengineering teams standardizing agent policysecurity and platform teams governing coding agentsmaintainers of agent plugins, skills and MCP servers

Deployment options

Local desktop app for macOS, Windows and LinuxPython package installed with pipxSelf hosted open source core (Apache 2.0)Guard Cloud, optional SaaS, region not stated

Integrations

Coding agents Claude Code, Codex, Cursor, GitHub Copilot CLI, Cline, Gemini CLI, OpenCode, Antigravity, Grok, Hermes, Kimi Code, Pi and Oh My Pi, OpenClaw and Z Code, through native hooks, managed MCP proxies and launch wrappers, with Devin and Paseo also in the support matrix. Guard Cloud on the Team plan connects Slack, Jira, PagerDuty, GitHub, email and HTTPS webhooks, with Notion in preview, and a routing step can ask a model through the customer's own OpenRouter account. The companion Plugin Scanner runs as a GitHub Action and writes SARIF.

In practice

A developer runs Claude Code and Codex side by side. Guard checks each shell command before it runs, blocks a recursive delete outside the project and pauses a read of cloud credentials until the developer approves it.

A coding agent tries to install a package published yesterday under a lookalike name. Guard checks it against supply chain advisories, blocks the install and writes a receipt explaining which rule stopped it.

A platform team sets a shared policy in Guard Cloud that laptops cannot loosen, routes high severity blocks to PagerDuty and Jira, and exports the evidence for its quarterly audit.

Agentic Index coverage score

9.0 / 14 capabilities · 64%

Integrations & Tool Calling Full

Guard connects to fifteen coding agents, among them Claude Code, Codex, Cursor, GitHub Copilot CLI, Cline, Gemini CLI, OpenCode, Hermes and Kimi Code. It reaches them through native hooks such as PreToolUse and PermissionRequest, managed MCP proxies that route a server's tool calls through Guard, and launch wrappers. How deep it goes depends on the events each agent exposes. On the Team plan, Guard Cloud connects to Slack, Jira, PagerDuty, GitHub, email and HTTPS webhooks, and review actions can be taken from Slack.

SourceHOL, github.com/hashgraph-online/hol-guard harness support matrix and hol.org/guard/featuresread 2026-11-17

Workflow Orchestration Partial

Every action takes one fixed path: a hook or wrapper catches it, a bundled Rust runtime checks it against policy, Guard allows, asks or blocks, and a receipt is written. Guard Cloud adds a routing studio where a team builds multistep graphs of triggers, enrichment, conditions, rate limits and delivery steps, tests them on a sample event and publishes them, with failed deliveries retried through an outbox. Routing only delivers notifications and records, and it never approves, blocks or changes what runs on a machine. Guard does not coordinate the agents it protects.

SourceHOL, hol.org/guard/docs/guard/routing.md and the Guard architecture guideread 2026-11-17

Knowledge Grounding & RAG Partial

An inventory of the plugins, skills, hooks and MCP servers on each machine is kept and can be exported as an AI bill of materials in JSON. Package installs are checked against supply chain intelligence, and HOL's Security Hub publishes advisories, CVEs, threat campaigns and monthly reports as RSS and JSON feeds, with premium feeds on Pro. HOL also runs a public plugin registry and a plugin security dataset. All of this is reference data that informs Guard's decisions, and there is no retrieval store over the customer's own knowledge for agents to draw on.

SourceHOL, github.com/hashgraph-online/hol-guard and hol.org/guard/llms.txtread 2026-11-17

Human Oversight & Guardrails Full

Ambiguous actions pause for a person. A held request can be resolved in the agent's own dialog where one exists, in a local approval center on the developer's machine, or from the terminal, and each approval takes a scope that runs from one artifact up to a publisher, workspace, agent or everything. One time approvals expire after 5 or 15 minutes, a timeout never counts as approval, and an optional password gate protects broad allows and history clears. On Codex, approving in the browser resumes the same thread. Team plans add a shared review queue with assignment.

SourceHOL, hol.org/guard/docs/guard/approval-center-and-audit.md and harness support matrixread 2026-11-17

Security, Identity & Governance Partial

Team workspaces carry members, roles and role based access, service principals for automation and API access, and named owners for machines, policies and cases. A managed restrictive control set lets an organization lock settings that local remembered decisions cannot loosen. Cloud commands use signed capabilities that are bound to one device or workspace, limited to exact operations and set to expire, and headless machines pair through an OAuth device code. HOL states no single sign on, no SOC 2 report, ISO certificate or trust center, and makes no compliance claims for Guard.

SourceHOL, hol.org/guard/features, local vs cloud guide and hol.org/guard/methodology/product-factsread 2026-11-17

Observability & Auditability Full

Every evaluation writes a local receipt with the reason, the policy action and the approval source. The audit trail keeps the pending request, its changed fields, the artifact that triggered it, the decision and the scope it was saved at. A policy check shows the effect, the winning rule and every candidate rule, so a block can be traced to its cause.

Guard Cloud adds searchable history across sessions and machines, incident summaries and evidence exports, and Team adds cases with an audit history of admin and review actions. History runs 7 days on the free plan and 30, 180 or 365 days in the cloud by plan.

SourceHOL, approval center and audit guide, policy as code guide and hol.org/pricing.md; hol.org/guard/docs/guard/approval-center-and-audit.mdread 2026-11-17

Memory & State Persistence Not documented

An always allow decision is saved against a command's hash and holds until someone removes it or the command changes. Paid plans sync this decision memory across sessions and devices, and Team shares reviewed policy decisions across a workspace. Guard applies these as rules the next time a matching action arrives. There is no memory layer for the agents Guard watches, and Guard runs no agent of its own that carries context from one run to the next.

SourceHOL, hol.org/guard/docs/guard/approval-center-and-audit.md and hol.org/guard/featuresread 2026-11-17

Deployment & Data Residency Full

The core runtime runs on the developer's own computer, installs without an account and keeps enforcing with no internet connection or while the cloud is down. History stays in a folder on the machine, nothing leaves it by default and file contents are never sent. Desktop builds cover macOS 13 or later on Apple silicon, Windows x64 and Linux x86_64, and the open source core can be self hosted. Guard Cloud is optional and receives receipts, redacted summaries and a fleet posture summary once a machine is paired. Its hosting region is not stated.

SourceHOL, hol.org/guard and the local vs cloud guideread 2026-11-17

Prebuilt Agents / Templates / Packs Partial

Starter policies ship with the product, a package firewall starter among them, along with policy presets and reviewed recipes that carry exact matchers, tests and downloadable hashes. Command rules come from an extension registry where core protections stay on and outside contributions need an explicit opt in, and the routing studio offers workflow templates. These are policy content for Guard itself, and HOL offers no prebuilt agents.

SourceHOL, policy as code guide, github.com/hashgraph-online/hol-guard and hol.org/guard/llms.txt; hol.org/guard/docs/guard/policy-as-code.mdread 2026-11-17

Triggers & Channel Coverage Full

Checks start on the agent's own events, with no prompt from a person. Hooks fire before each tool call and on permission requests, and on Copilot CLI, Kimi Code, Z Code, Devin, Cline and Pi they also fire when a prompt is submitted, so shell commands, file reads, MCP calls and package installs are evaluated before they execute. Launch interception reviews new or changed MCP servers, plugins, skills and hooks before an agent starts. Guard Cloud flags MCP and skill drift and sends digests and matched advisories, and routing workflows start on synced events such as a new triage item.

SourceHOL, harness support matrix and hol.org/guard/features; github.com/hashgraph-online/hol-guard/blob/main/docs/guard/harness-support.mdread 2026-11-17

Model Flexibility & Routing Partial

Allow, ask and block decisions come from a policy engine, and no model takes part in enforcement. A model appears in one place, a Guard Cloud routing step that sends an investigation event through the customer's own OpenRouter account and records the answer. The customer sets the model on that step, with a free OpenRouter model as the default and paid models billed by OpenRouter. The answer never changes a Guard decision.

SourceHOL, hol.org/guard/docs/guard/integrations/openrouter.mdread 2026-11-17

APIs / SDKs / MCP Extensibility Partial

The core is open source under Apache 2.0 on GitHub, written in Python with a Rust runtime and published on PyPI. Policies are text files in a published GuardPolicy format that is still in alpha, and the command line covers status, approvals, receipts, inventory, policy tests and pairing. Team plans add service principals for API access, and HOL lists an OAuth protected MCP endpoint for workspace and receipt tools. No API reference or tool list for Guard Cloud is published.

SourceHOL, github.com/hashgraph-online/hol-guard, hol.org/llms.txt and hol.org/guard/featuresread 2026-11-17

Testing, Debugging & Optimization Full

Policy changes can be tested before they take effect. Test cases live in the policy file, each describing an event and the effect expected, and the test command reports pass, fail or ambiguous, with exit codes that gate a CI pipeline and a diff check against the live policy. New policies can run in observe mode, recording decisions on real agent activity without stopping anything, before they move to enforce. Imports preview by default, and a saved export rolls a policy back. The companion Plugin Scanner also runs as a GitHub Action quality gate on agent plugins.

SourceHOL, hol.org/guard/docs/guard/policy-as-code.mdread 2026-11-17

Browser / Computer-use Not documented

Guard watches what the customer's coding agents do, such as shell commands, file reads, package installs and MCP calls, and it drives no browser, desktop or remote computer itself. Browser actions an agent takes through MCP can be approved for 15 minutes, 1 hour or 5 hours at a time, and the agent doing the browsing is the customer's own.

SourceHOL, hol.org/guard/docs/guard/approval-center-and-audit.mdread 2026-11-17

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Free on one machine. Solo is $4.99 a month, Pro $15 a month and Team $30 a seat a month, with Enterprise by contract.

Solo and Pro are billed per user and Team per seat, monthly or annually.

Free tierTrial available

Included quota

Solo covers 2 cloud connected devices, 30 days of history and 1 GB of storage. Pro covers 5 devices, 180 days, 5 GB and 500 starter credits. Team covers up to 25 devices, 365 days, 25 GB plus 5 GB a seat and 2,000 credits. Enterprise covers up to 500 devices.

What is public

HOL Guard's paid plans run through Guard Cloud. Solo is $4.99 a month, Pro is $15 a month and Team is $30 a seat a month. Each plan gives two months free when billed annually. Enterprise is priced by contract.

Billing mechanics

Each plan caps the number of cloud connected devices. Credits sit in a separate balance for automation work such as API scanning and policy checks, and local protection never draws on them.

Cost watchouts

Cloud automation such as API scanning draws on credits, which are bought apart from the plan.

Variable cost rationale

Each plan is a flat price per user or seat. Extra spend comes from added seats, a larger plan for more devices, or credits for cloud automation.

Overage / add-ons

A device over the plan limit keeps local protection and stops syncing to the cloud. More credits are bought from the billing screen.

Sales call required

Mixed (some tiers require a call)

Free / trial

Guard Local is free forever on one machine, with no account, local policy decisions, receipts and 7 days of history. Solo, Pro and Team start with a 3 day trial.

Lowest paid plan

Solo, $4.99 a month

Commercial notes

Solo, Pro and Team are bought online, and the open source core keeps working without a paid plan. Enterprise starts with a security review covering identity and access, redacted evidence export, custom threat feeds, volume pricing, support and self hosted or VPC deployment. Its terms are set for each deployment.

Key ambiguities

Pro and Team mention optional automation wallets and overage backed automation, with no figures given.

Cancellation / refund

Paid plans renew monthly or annually. When a trial ends or a payment fails, local protection keeps working and cloud features pause.

Missing data

HOL does not publish the price per credit, its refund terms or an Enterprise price.

Agentic Index verified 2026-11-17

HOL Guard pricing, plans and cost →

Alternatives to HOL Guard

Other products that do the work HOL Guard does, and what each one is.

HOL Guard alternatives and competitors →

Head to head

Contact us

Found a vendor we missed? Have feedback on the index? We'd love to hear from you.