Comparison

Self-hostedAI agent platforms,compared.

6 platforms for organizations that want to keep agents, data and credentials in their own environment, compared on how they deploy, how they are sold, whether people can approve agent actions, how they are licensed and which security controls they document.

How this comparison works

Verified
October 13, 2026
Cited claims
42

Disclosure: Beag Labs makes Papyrus, one of the platforms below. Papyrus is cited from our own pages and labelled first-party, and the alternatives are given the same evidence rules and a fair account of where they may fit better. See the methodology.

Comparison table

Filter by how you need to deploy and buy.

Each cell shows its evidence status and links to the Claim Ledger record that holds the source. Open “sources checked” on a Not documented cell to see where we looked.

Known limitation: Microsoft Marketplace pages did not load for our automated checks. Marketplace cells therefore rest on vendor pages, and a Not documented Marketplace cell means we found no listing in the sources we could read. Confirm offer details on the listing itself before buying.

Showing 6 of 6

Self-hosted AI agent platforms compared by attribute, with a source link on every verified cell
PlatformDeployment modelAzure Marketplace offerHuman approvalLicensingSecurity controlsIntended buyerPublic pricing
PapyrusBeag LabsMade by Beag LabsCustomer-hosted agentic workspace for document automation, internal app creation and legacy system enablement.
Verified

Customer-hosted: a Papyrus VM in the customer's Azure subscription, or on-premises, restricted or disconnected environments.

Source: BL-CLM-19 ↗
Verified

Azure Marketplace private offer that deploys the Papyrus VM into the customer's subscription.

Source: BL-CLM-20 ↗
Verified

Governed execution: consequential actions are reviewed and work stays visible as it progresses.

Source: BL-CLM-21 ↗
Verified

Commercial: a signed organization license (offline) or a customer-specific Azure private offer.

Source: BL-CLM-22 ↗
Not documented

Not found in checked official pages.

1 source checked
Verified

Organizations that need back office automation on their own infrastructure, in business and government settings.

Source: BL-CLM-24 ↗
Verified

No public price card; commercial terms are scoped to the deployment.

Source: BL-CLM-25 ↗
Corral AI (identified as 'Corral Sovereign AI')CorralCorral AI is described by its vendor as an AI foundation deployed from Azure Marketplace into the customer's own Azure tenant, where agents are built and governed.
Verified

Deploys into the customer's own Azure tenant and subscription; AWS and GCP listed as coming 2026.

Source: BL-CLM-27 ↗
Verified

Vendor site links a 'Corral AI managed app' Azure Marketplace listing (azure-application URL path).

Source: BL-CLM-28 ↗
Partially supported

Admin approval gates the agent publish flow, and high-risk actions need a human after untrusted content; other actions are not described.

Source: BL-CLM-29 ↗
Not documented

Not found in checked official pages.

2 sources checked
Verified

Documented: on-tenant residency, private endpoints/VNet, immutable audit trail, content filtering/PII detection, identity-provider governance.

Source: BL-CLM-30 ↗
Verified

Lists Ops & Product Teams, IT & Security Teams, Developers, Service Providers/SIs, Product Companies.

Source: BL-CLM-31 ↗
Verified

No prices listed; platform subscription plus Azure infra costs via Marketplace, counting toward MACC; Enterprise plan via 'Talk to us'.

Source: BL-CLM-32 ↗
HeartBeatAgentsMosaic SingularityHeartBeatAgents is a self-hosted platform for building and running autonomous AI agents that work on standing orders with local memory.
Verified

Self-hosted: desktop download (macOS DMG) on vendor site; AWS listing ships an AMI run in buyer's AWS account.

Source: BL-CLM-34 ↗
Not documented

No Azure Marketplace listing found; only an AWS Marketplace AMI listing was found.

3 sources checked
Partially supported

Only simple 'Ask User' and 'Present Choices' skills let an agent request approval or a choice; no broader approval policy is described.

Source: BL-CLM-35 ↗
Verified

AWS listing: EULA applies; community edition with one worker, more workers via separately purchased license.

Source: BL-CLM-36 ↗
Verified

Documented: AES-256 at rest, TLS 1.2+, per-organization container folder scoping, token-authenticated websockets, token scrubbing.

Source: BL-CLM-37 ↗
Verified

AWS listing targets organizations wanting production-ready self-hosted agentic AI infrastructure.

Source: BL-CLM-38 ↗
Verified

No prices listed; AWS listing says community edition with one worker, billing external via vendor license.

Source: BL-CLM-39 ↗
OpenClawOpenClaw Foundation (independent nonprofit, per repo README)OpenClaw is an open-source assistant/gateway that runs on your own machine or server and connects chat apps to AI agents (repo README; docs.openclaw.ai).
Verified

Runs on your own laptop or server; install via script, npm, Docker, Nix or source; remote access via SSH/Tailscale. README says no hosted service.

Source: BL-CLM-41 ↗
Not documented

No official Azure/Microsoft Marketplace listing found in first-party pages. A web search surfaced third-party publisher listings (not fetched; unverified).

3 sources checked
Partially supported

Exec approvals (deny, allowlist, ask, auto, full) gate host command execution only and can be switched off; not a general approval layer.

Source: BL-CLM-42 ↗
Verified

MIT License, copyright OpenClaw Foundation (2026).

Source: BL-CLM-43 ↗
Partially supported

Gateway auth, DM pairing and allowlists, optional sandboxing (off by default) and an audit CLI; documented as a single-trusted-operator model, not a multi-tenant boundary.

Source: BL-CLM-44 ↗
Verified

Individuals wanting a personal assistant, teams sharing a gateway, and enterprise/security teams hardening via configuration, not a separate edition.

Source: BL-CLM-45 ↗
Verified

README says no paid tier, hosted service, or token. Model API and hosting costs are borne by the user.

Source: BL-CLM-46 ↗
OpenHandsOpenHands (company name not stated on pages fetched)OpenHands is an open-source platform for software-engineering agents, with a GUI, CLI, SDK and enterprise control plane (openhands.dev).
Verified

Local/self-hosted (npm, Docker), plus OpenHands Cloud (managed) and OpenHands Enterprise (licensed self-host in your VPC or managed).

Source: BL-CLM-48 ↗
Not documented

No Azure/Microsoft Marketplace listing found for OpenHands in first-party pages or search.

4 sources checked
Partially supported

SDK confirmation policies (always, never, risky) can require approval of actions; documented for the SDK, and one call path bypasses them.

Source: BL-CLM-49 ↗
Verified

Root LICENSE of OpenHands/OpenHands is MIT (OpenHands contributors, 2025). Docs say each repository has its own license.

Source: BL-CLM-50 ↗
Verified

Docs name: Docker sandbox isolation for the agent server, security analyzers with risk levels, Enterprise SAML/SSO, RBAC and access controls (Cloud).

Source: BL-CLM-51 ↗
Verified

Developers and engineering teams, including enterprises wanting control over models, data, access and auditability.

Source: BL-CLM-52 ↗
Verified

Local open source: free. SaaS Individual: free, 10 daily conversations. Enterprise: custom pricing, contact sales.

Source: BL-CLM-53 ↗
Websoft9Websoft9Websoft9 is a self-hosted PaaS and app store for deploying and managing 300+ open source apps on one server.
Verified

Self-hosted on your own server, running as Docker containers.

Source: BL-CLM-55 ↗
Verified

Docs say Websoft9 is an Azure Marketplace partner; products are VM images deployed to an Azure VM.

Source: BL-CLM-56 ↗
Not documented

No approval or human-in-the-loop controls documented for agent actions.

2 sources checked
Verified

Repository states LGPL-3.0 and homepage says 100% open source; republishing images to cloud marketplaces without authorization is prohibited.

Source: BL-CLM-57 ↗
Verified

Multi-user accounts with role-based access; SSL certificates via Let's Encrypt; apps isolated through Docker.

Source: BL-CLM-58 ↗
Verified

Cloud architects, IT teams and SMBs without dedicated IT, plus software vendors, integrators and cloud providers.

Source: BL-CLM-59 ↗
Verified

Software plans: Free 0, Starter CNY 20/mo, Standard CNY 99/mo (per vendor pricing page); Azure image billing not described.

Source: BL-CLM-60 ↗

Where each one may fit

No platform wins every comparison.

Fit statements below are derived only from cells in the table, and each lists the cells it rests on. They are not rankings.

Papyrus

Made by Beag Labs

Customer-hosted agentic workspace for document automation, internal app creation and legacy system enablement.

May fit well when

  • Teams that must keep agent execution inside their own Azure subscription or a disconnected environment.Basis:Deployment model (BL-CLM-19)
  • Buyers who need a published, machine-readable control mapping to review during procurement.Basis:Security controls

May fit less well when

  • Teams that want a free, open-source licence or public per-seat pricing to evaluate without talking to sales.Basis:Licensing (BL-CLM-22)

Corral AI is described by its vendor as an AI foundation deployed from Azure Marketplace into the customer's own Azure tenant, where agents are built and governed.

May fit well when

May fit less well when

HeartBeatAgents is a self-hosted platform for building and running autonomous AI agents that work on standing orders with local memory.

May fit well when

May fit less well when

  • Buyers who need to purchase through Azure Marketplace: we found no Azure listing in the sources checked (an AWS Marketplace listing exists).Basis:Azure Marketplace offer

OpenClaw is an open-source assistant/gateway that runs on your own machine or server and connects chat apps to AI agents (repo README; docs.openclaw.ai).

May fit well when

May fit less well when

OpenHands is an open-source platform for software-engineering agents, with a GUI, CLI, SDK and enterprise control plane (openhands.dev).

May fit well when

May fit less well when

Websoft9 is a self-hosted PaaS and app store for deploying and managing 300+ open source apps on one server.

May fit well when

May fit less well when

  • Buyers needing documented approval gates or audit logging for agent actions; these are not documented.Basis:Human approvalSecurity controls (BL-CLM-58)

Methodology and disclosure

How these comparisons are built and checked.

We disclose our interest

Beag Labs makes Papyrus, one of the platforms compared here. Papyrus cells cite our own pages and are labelled first-party. Read every comparison with that in mind.

Every statement is cited

Each Verified or Not supported cell links to a record in the Claim Ledger with the source URL, where on the page it appears, the date we checked it, approved wording, and what it does not prove.

Documentation, not hands-on testing

These comparisons are built from vendor documentation, official listings and repositories. Beag Labs has not run hands-on tests of the compared products. We classify what is documented; we do not score or rank quality, and nothing here is a benchmark.

Missing documentation is not a missing feature

If we cannot find a capability in official sources we say Not documented and list what we checked. We only say Not supported or Partially supported when a source itself says so.

Verified
An official source we fetched states it. The cell links to the claim record.
Partially supported
An official source documents the capability but states a limit: one surface, a subset of actions, or an explicit exclusion. The cell says what the limit is.
Not documented
We checked official sources and found nothing either way. This is not evidence the capability is absent.
Not supported
An official source explicitly says the capability is absent or unsupported.

Last verified October 13, 2026. Vendors change products and pricing; treat these pages as a starting point and confirm anything decisive with the vendor. If we got something wrong or a vendor has newer documentation, write to james@beaglabs.com and we will re-check the claim and update the ledger. Browse every record in the Claim Ledger.

Questions about this page

Is this comparison independent?

No. Beag Labs makes Papyrus, one of the platforms compared. We disclose that on every comparison, label Papyrus cells as first-party, and apply the same evidence rules to every platform, including a fair account of where each alternative may fit better.

Did you test these products?

No. These comparisons are built from vendor documentation, official listings and public repositories. We have not run hands-on tests or benchmarks of the compared products, and we say so rather than imply otherwise.

What does Not documented mean?

We checked the official sources listed on the cell and found nothing either way. That is not evidence the capability is missing. We only mark a capability Not supported when an official source explicitly says so.

Why no scores or star ratings?

Scores would imply testing we have not done. Our evidence is documentation, usually one page per attribute, so we classify what is documented, partially supported, unsupported or undocumented, and state the limits of each source. Hands-on testing would be needed before rating quality.

What does Partially supported mean?

An official source documents the capability but states a limit, such as covering one surface, a subset of actions, or having an explicit exclusion. The cell spells out the limit and links to the source.

Where does each statement come from?

Every Verified or Not supported cell links to a record in the Claim Ledger with the source URL, where it appears on the page, the date it was checked, the wording we may use and what the source does not prove.

A vendor thinks something here is wrong. What happens?

Write to james@beaglabs.com. We re-check the source, correct the cell and ledger record if needed, and update the verification date.