OPEN SOURCE
We open-source the parts you would otherwise have to take on faith.
Four projects: the data factory that decides what a model may learn from, the gateway that controls what an agent may call, the registry that records which model is in production, and the tooling that labels the data. Each one makes a claim about governance — and a governance claim you cannot read is a claim you cannot check.
Why we are giving this away.
AI is being built faster than it can be governed, and the gap is widest exactly where the stakes are highest — banking, defence, healthcare, the public sector — where "the model is probably fine" was never an acceptable answer. We do not think the answer is to slow AI down. We think it is to make AI verifiable: you should be able to prove what a model learned from, what an agent is allowed to do, and what is actually running in production — not take it on faith, and not take our word for it.
Governance you cannot inspect is only a promise, and for institutions that carry real-world risk a promise is not enough. So the components that make these proofs possible should not be locked inside a vendor. If we ask you to trust how we govern AI, you should be able to read the code that enforces it — which is why we are opening our control points to the community, starting with Shuddhi.
This is AgentAnywhere contributing back. Clean, regulated, responsible AI is not a brake on progress — it is the foundation that lets AI be deployed where it actually matters. We are building that foundation in the open, and we invite you to build on it with us.
Why these four, and not everything.
Opening a codebase is not, by itself, a virtue. The question worth answering is which components a customer, a regulator or a sceptic would need to read in order to believe what we say — and those are the ones that should be open.
All four sit at a control point. Shuddhi decides what enters a training corpus. The Universal Gateway decides which call an agent is allowed to make. Model Hub records which model is actually serving. The Annotator produces the labels the whole chain rests on. If any of these quietly did something other than what we describe, every downstream assurance would be worthless — which is precisely why they are the ones we are opening.
The rest of the platform — the sovereign deployment topology, the signed audit chain, the regulatory mappings, SSO and multi-tenancy, the support commitments — is commercial. That boundary is drawn at where the work is, not at where the leverage is: those components take a paid team to keep current with regulation, and none of them is something you would need to read our source to trust.
The four projects, and where each one stands.
Shuddhi — training-data provenance
Provenance gating, deduplication, language ID, quality and domain balancing, PII redaction, contamination screening — sealed with a build receipt a training run can cite, plus a draft EU AI Act Article 53(1)(d) training-content summary. Licence: Apache-2.0. Status: PUBLISHED — live at github.com/agentanywhere/shuddhi. Run it in one command: docker run --rm ghcr.io/agentanywhere/shuddhi:1.2.0 demo.
Agent Universal Gateway — call control
The dual-protocol MCP + A2A control plane every agent call passes through, with the policy evaluation that decides whether a call proceeds — nine protocols from one surface. Licence: Apache-2.0. Status: release-ready. The open edition is extracted and sanitised, five launch-blocking defects were fixed with regression tests that fail on the pre-fix code, and its suite runs 648 green on a clean machine.
Model Hub — model records
Model identity, versioning, checksums and lineage — the record of which model is in production and what it was built from. Licence: not yet fixed. Status: earliest stage of the four; scoping, with no code separated out yet.
Annotator — labelling tooling
The annotation and review tooling that produces the labelled data underneath the rest of the chain. Licence: Apache-2.0. Status: contribution and security policies complete; awaiting publication.
What is public today.
Being precise about this matters more than making the page look busy. Shuddhi is now public — the first of the four to ship: github.com/agentanywhere/shuddhi, Apache-2.0, with a prebuilt image you can run in one command. The Universal Gateway is release-ready — built, licensed, tested and sanitised, waiting only on the decision to publish — with the Annotator close behind. Model Hub is the earliest of the four and has no code separated out yet.
We would rather this page understate where we are than have you discover the gap yourself. Shuddhi is live today; as each remaining repository goes public, this page will name it and link to it. Everything not yet linked is a commitment rather than a download.
Where the boundary sits.
The same line is drawn in all four projects: the mechanism and the proof are open; running it as an organisation is commercial. Stated concretely so it can be held against us.
Open source (Apache-2.0)
- The engine, and every filter, check and policy primitive shipped in it
- Local verification — receipts and reproducible build hashes you can recompute yourself
- The refusal rules, readable in full, including the refusal of customer data
- The plugin and extension API, so commercial features attach through the same seam anyone else can use
- No licence fee, no user cap, and no telemetry in the open engines
Commercial
- Hosted registries and the signed attestation chain (TrustReceipt)
- Team review workflow — routing, approvals, sign-off and who-saw-what
- Cross-build reporting and analytics
- Sovereign and air-gapped deployment, multi-region, disaster recovery
- Enterprise SSO, RBAC, multi-tenancy
- Compliance evidence packs, sector rule sets, 24x7 support and SLAs
Permanently open.
What ships as open source stays open source. The engines, their built-in filters and checks, local receipt generation and the refusal rules stay Apache-2.0 and will not move behind a proprietary licence later. Commercial offerings extend these projects through the public extension API and through hosted services; they do not remove capability from the open engine, and an existing open feature will not be repackaged as a paid one.
This is a commitment we are making because the alternative is well documented. Several open-core companies relicensed after adoption, and what they lost was not revenue but the willingness of anyone to build on them again. The guarantee is worth more to us than the option.
The discipline that keeps it honest: if a commercial feature requires a change to an open engine, that change is made in the open repository. It is an extension-API improvement, and it prevents the open engine from being quietly special-cased for paid features.
FAQ
Frequently asked questions.
- Which AgentAnywhere projects are open source?
- Four: Shuddhi (training-data provenance), the Agent Universal Gateway (agent call control), Model Hub (model records and lineage), and the Annotator (labelling tooling). Shuddhi and the Annotator are Apache-2.0; the Gateway and Model Hub will follow under the same licence. Shuddhi is the first release. The wider AgentAnywhere platform is commercial.
- Can I download them today?
- Shuddhi, yes — it is public at github.com/agentanywhere/shuddhi under Apache-2.0, and you can run it in one command: docker run --rm ghcr.io/agentanywhere/shuddhi:1.2.0 demo. The Universal Gateway and Annotator are release-ready and follow; Model Hub is earliest-stage. This page names and links each repository as it goes public.
- Is AgentAnywhere Core open source?
- No. AgentAnywhere Core is not part of the open-source programme. The programme is the four projects named above: Shuddhi, the Agent Universal Gateway, Model Hub, and the Annotator.
- What licence do you use, and can I use it commercially?
- Apache-2.0 across the programme. It permits commercial use, modification and redistribution, imposes no copyleft obligation on your own code, and includes an express patent grant — which is the main practical reason we prefer it to MIT. There is no licence fee and no user cap. Shuddhi and the Annotator carry the licence today; the Gateway and Model Hub will ship under it.
- Will you relicense later, or move open features behind a paywall?
- No. What ships as open source keeps its open licence: the engines, their built-in filters and checks, local receipt generation and the refusal rules. Commercial offerings extend these projects through the public extension API and through hosted services rather than by removing capability from the open engine.
- What is commercial, then?
- Hosted registries and the signed attestation chain, team review workflow with approvals and sign-off, cross-build reporting, sovereign and air-gapped deployment, enterprise SSO, RBAC and multi-tenancy, compliance evidence packs and sector rule sets, and 24x7 support with SLAs. The engine and the proof are free; running it as an organisation is paid.
- Why open-source governance components at all?
- Because a governance claim you cannot read is a claim you cannot check. These four sit at control points — what a model may learn from, what an agent may call, which model is serving, and how the data was labelled. Their value depends on being inspectable, which makes them worth more open than closed.
Tell us which one you want first.
If one of the four would be useful to you, say so — what you would run it on and what you need it to prove shapes the order they go out in, and we will tell you when the repository is public. If you are evaluating the commercial platform on top, that is a different conversation and our team is the right path.