Now in Private Beta
Private AI agents on infrastructure you control.
Kovalent deploys agents across your cloud, edge devices, and private nodes, linked by an encrypted mesh. The model, its knowledge, and your data stay on your nodes; Kovalent runs only the coordination that connects them.
Who it is for
Built for the work you cannot hand over.
Most teams have some work they cannot put through a shared AI platform. These are the constraints we hear most often, and what the platform does about each one.
Regulated data
Your data sits under a regulator, and "our vendor promised not to" is not an answer you can file.The control plane has no route into your node's store. It sends intents, not queries, so what a reviewer is shown is an architecture rather than an assurance.Commitments you made
You have told your customers their documents will not be handed to a third-party AI provider.Your knowledge base and your chat history stay on your own node, and the retrieval that reads them runs there too. The commitment stays true without anyone having to remember it.Data nobody will pool
Two teams each need AI over data that neither of them will put into one shared index.Each keeps its own node and its own index. A question can be answered across both without the documents behind it ever being combined in one place.Systems on your network
The things your agents have to reach are inside your network, not out on the public internet.Nodes run where those systems already are: your cloud account, your cluster, or a machine on your desk. An encrypted mesh joins them without exposing any of them.Already using a shared AI platform? Keep it, and keep the models your team likes. Kovalent routes to the same frontier models, and to open ones you run yourself, while your knowledge base and chat history stay on a node you control.
Model support
The model is a part you swap, not a platform you join.
Where the model runs is a setting on the node, not a decision about your architecture. There are three places it can be, and moving between them does not touch anything else.
A hosted frontier modelYour node brokers the call out to a hosted model, so a small node still answers with a frontier one. Each tenant calls under its own cloud identity rather than a pooled platform key.
Leaves the nodeThe prompt and its context
An open model, in your nodeThe model runs on the node itself, reachable only on its own loopback. The control plane never sees the call, only the finished answer, so no third-party inference sits in the path.
Leaves the nodeNothing
A runtime you already runPoint a node at Ollama, LM Studio, vLLM, or llama-server on your own machine. The CLI finds whichever is answering and uses the model you already pulled.
Leaves the nodeNothing
Frontier and open weights alike
A node set to generate in-pod does not quietly fall back to a hosted model when its own runtime is missing. It refuses, and says so. Silent fallback is how a private deployment stops being one without anybody noticing.
How it works
Kovalent runs the control plane. You own your nodes.
An agent is really several parts. There is the model, its knowledge base, and the runtime that puts them to work. Then there is the coordination that ties all three together. Kovalent runs only the coordination. The model, its knowledge, and the runtime stay on nodes you own. Computation stays with you, and only the coordination is managed.
Register a NodeActive
Connect a machine, cloud instance, or private cluster as a cryptographically identified peer.Attach KnowledgeActive
Store embeddings and retrieval context on your node, next to the data source, instead of in a shared vendor runtime.Route Over MeshActive
Send work over P2P WireGuard paths without exposing raw data through the control plane.Run AgentsActive
Execute reasoning, retrieval, and tool workflows inside isolated node runtimes.Audit MovementIn Beta
Stream lifecycle metadata so operators can see routing and collaboration without inspecting payloads.Extend SafelyNext
Bring private APIs, storage, and future tool runtimes into the mesh behind customer-controlled policy.Deployment options
The quickest node to start is one you already own.
The same node runs in all four places. What changes is whose machine it sits on, and you pick that per node rather than once for the whole account. Nothing here asks you to trade being close to your data for being quick to start.
Fastest start
Your own machineOne command turns a laptop or a server you already own into a node. Sign in and it appears in your dashboard beside the rest, reporting its own health, metrics, and logs.- Machine
- Yours
- Best when
- You want a node answering today, on hardware you already have.
Bring your own network
Your own networkNodes attach to the network you already run, inside your existing perimeter and under your existing rules.- Machine
- Yours, on your network
- Best when
- The systems being reached are not on the public internet.
Single tenant
Dedicated nodeA machine provisioned for you alone, in a region you choose at the time.- Machine
- Yours, run by us
- Best when
- Production work, or a jurisdiction you have to name.
Nothing to run
Managed cloudA pod on our cluster, provisioned in minutes, with no hardware of yours involved and nothing to keep running.- Machine
- Ours
- Best when
- You would rather not operate a machine at all.
Wherever a node runs, its knowledge base and chat history live on it, and the control plane reaches it with intents rather than queries. See what each tier includes.
Coordinated centrally. Connected directly.
Kovalent handles coordination, not computation. You choose which nodes run your AI. Kovalent manages the identity and routing that let those nodes find each other, while inference and data stay on infrastructure you control. The nodes connect directly over an encrypted private network built on standard WireGuard, so Kovalent coordinates the mesh without ever seeing the data that moves across it.
The control plane (coordination)
We handle SSO identity, Tailscale ACLs, and node discovery. The control plane issues routing directives but never touches your raw data streams.Your nodes (computation)
Agents run on isolated hardware (such as dedicated EC2 instances). Connecting a new peer via knaix login initiates direct NAT traversal for low-latency inference.Zero-Trust Architecture
Every node is assigned a cryptographic identity. A deny-by-default firewall ensures that nodes only accept traffic from authorized peers within the mesh.NEW - DISTRIBUTED INTELLIGENCE
From a secure mesh to a collaborative one.
Your nodes don't just share a network. They share intelligence. Each node keeps its own model and knowledge. Inside a Tenant Mesh, agents discover one another, route each question to whichever node owns the domain, and synthesize answers together. No raw data ever leaves a node; only the answer travels.
Relevance-based discovery
Every node publishes a domain manifest. The orchestrating node scores its peers and delegates only to the best match, never a blind broadcast.Map-reduce consensus
Premium nodes synthesize answers and return a compressed result, so raw chunks never cross the mesh and the token cost of querying it drops.Real-time mesh telemetry
Watch every hop as it happens. Peer selection, federated search, and synthesis stream live to your console. Metadata only; we stay blind to your data.COMING SOON TO TEAM & ENTERPRISE
Deploy your fleet.
A general-purpose model handles most of what your team asks. Specialized work deserves specialized models. Pick them for image generation, document extraction, and storage optimization, then deploy them across your nodes. Swap or upgrade each one as your needs change. Every fleet model runs inside your boundary, on your nodes or under your own cloud identity, never through a shared platform runtime.
Curated model catalog
Every catalog entry clears the same provenance and license review as the models the platform ships with today. The right tool for each workload, vetted before it reaches your mesh.Swap and upgrade in place
Pin a version per workload, upgrade on your schedule, and roll back in one step. Your agents keep working while the models underneath them change.Specialist routing, general fallback
Requests route to the specialist you deployed for that job. Everything else falls back to your general-purpose model, so the fleet grows at your pace without breaking anything.NEW
The Knaix CLI/nɑks/ Rhymes with "knocks"
Deploy and manage nodes directly from your terminal, or run the whole stack on your own machine with your own local model. Provision securely, connect instantly, and let the Mesh handle zero-trust networking without complex configuration.
Developer First
- Local Runtime
Run the whole stack withknaix local, no account. Point it at your own Ollama or LM Studio model. - Node Auth
SSO login for your machine. Your key, your private mesh. - Node Bridge
Hardware becomes a trusted peer for local inference. - Skill Orchestration
Trigger reasoning skills on your isolated local data. - Local, in your Dashboard
Connect a local node withknaix local connectand it appears in the Kovalent Dashboard with its metrics and logs, next to hosted nodes. No paid plan needed.
Secure by Design
- Zero-Trust Network
No inbound firewall rules. The CLI tunnels out securely. - Ephemeral Operations
Logs and sessions do not leak. Nothing caches centrally.
# Install the Knaix CLI
$ brew install kovalentai/tap/knaix
# Run the whole stack locally, no account
✓ Node 'local' is ready: store, embedder and reranker, on your machine.
No model yet, so answers come from the labeled mock.
Info: Probing the usual model ports (Ollama, LM Studio, vLLM, llama-server)...
✓ Ollama at http://localhost:11434 → answering with llama3.1:8b.
# Or sign in to provision a hosted node
Info: Starting Knaix SSO Login...
Opening browser: https://app.kovalentai.com/cli-auth...
✓ Successfully logged in!
Info: Synchronizing with private mesh..
⠋ Requesting compute...
⠼ [BUSY] Booting kernel for claude-e146c33e46...
✓ Node claude-e146c33e46 provisioned successfully.
● Memory | Node: claude-e146c33e46
Use `knaix memory --file <filename>` to read a file.
_knaix_durable_memory.md
_knaix_ephemeral_log.md
● Knaix AI Session: claude-e146c33e46 (Type '/exit' to end)
knaix [claude-e146c33e46]> /remember Create an ingress gateway routing rule.
Intent recognized: Create an ingress gateway routing rule.
✓ Explicit memory securely stored and available across sessions.
Security & governance
Ready for your security review.
Every deployment starts from one guarantee: the control plane coordinates identity and routing, and never sees your prompts, your documents, or your model traffic. Here are the controls behind that, what runs today, and what each tier adds.
Available now
Content-blind control plane
Kovalent coordinates identity, routing, and lifecycle. It never sees your prompts, documents, or model traffic.
Available now
Per-tenant inference identity
Model calls run under your own cloud identity and quota, never a shared platform key.
Available now
Encrypted peer transport
Nodes connect directly over WireGuard (Curve25519). Traffic flows peer-to-peer, not relayed through us.
Available now
Deny-by-default mesh
Every node carries a cryptographic identity and accepts traffic only from peers you have authorized.
Personal+
Your knowledge stays on your node
Documents and chat history live in your own node, and the ranking that reads them runs there too. Data residency by architecture, not by policy.
Professional+
Dedicated hardware isolation
Professional and above run on dedicated single-tenant instances with kernel-level isolation, not shared pods.
Professional+
Bring your own network
Map nodes into your existing Tailnet under your own identity provider and ACLs, or peer them into your VPC.
Enterprise
Air-gapped operation
Enterprise nodes can run inference locally with no external model calls at all, for regulated or disconnected environments.
Planned
Audit, SSO, and compliance profiles
Audit logs, SSO and team administration, SIEM export, and regional data-residency profiles that attest which jurisdiction a workload ran in, for governed Enterprise deployments.
Preparing a security review? See what runs on each tier, see what is coming, or talk to us about compliance.
Positioning
How Kovalent compares.
Different approaches to private AI solve different problems. Here is where Kovalent fits and where it does not.
vs. vendor AI suites
ChatGPT Enterprise, Claude for Work, Copilot
Their privacy is contractual: your data stays in their infrastructure under their terms. Kovalent's privacy is architectural: per-tenant identity, a control plane that cannot see payloads, and nodes that run under your own cloud credentials. They centralize; we federate custody.
vs. central RAG platforms
Glean, enterprise search tools
Excellent when every team is willing to share one index. Kovalent starts where that willingness ends: teams that refuse the data lake, subsidiaries with residency constraints, and organizations where combining indexes creates a liability.
vs. self-hosted open source
Open WebUI, AnythingLLM, llama.cpp
Full sovereignty, zero product. No managed provisioning, no encrypted mesh, no cross-node federation, and a permanent ops burden. We deliver the same sovereignty posture with orchestration, networking, and retrieval built in.
vs. assembling it on AWS
Bedrock Knowledge Bases, SageMaker, glue code
A capable team can assemble the retrieval layer. They cannot cheaply assemble the federated mesh, the containment boundary, the per-tenant identity model, or the sovereign inference fallback. That integration surface is what Kovalent is.
Roadmap
What ships today, and what comes next.
24 capabilities run in production today, and 32 more are committed and dated. No feature on this page is aspirational: every one of them is tracked work.
Last reviewed against the platform on 25 July 2026.
Available now24In beta3Q3 20267Q4 202610202712
Running in production today
Infrastructure and networking
Where your AI runs, and how those machines reach each other.
Private node orchestration
Provision a node as a shared pod, a private pod, or dedicated EC2, and the control plane handles its lifecycle from there.
One runtime on every path
The same Node Runtime serves your node whether it runs on Kubernetes, on a dedicated instance, or on your own laptop.
Knowledge and retrieval
Turning your documents into answers you can check.
Inference
The model itself, and where the thinking happens.
Console, CLI, and API
The ways you drive it: a browser, a terminal, or your own code.
Knaix CLI
Sign in, provision and manage nodes, or run a node on your own machine with no account at all. Open source, Apache-2.0, on macOS, Linux and Windows.
knaix doctor and knaix benchDiagnose a node that is misbehaving, and measure what yours actually delivers.
Node console with addressable surfaces
Overview, knowledge, chat, metrics, usage, logs, and settings each have their own URL, so any of them can be bookmarked or sent to a teammate.
Infrastructure observability
Live metrics and logs for each node, read from the node itself.
Security and governance
Proving that what we claim about your data is actually true.
Signed releases you can check yourself
Every release is signed by the workflow that built it, so there is no signing key to steal.
knaix verify checks the binary you are running: its checksum, its signature, and how it was built.