Documentation

Hanzo OS · Since 2014

Everything your company needs. Built together.

The complete operating system for the modern company.

Frontier intelligence, AI coworkers, the applications a company runs on, secure execution, company data, full-stack observability and Kubernetes-native infrastructure — designed together rather than assembled.

Run Hanzo OS on Hanzo Cloud or your own infrastructure.

Try HanzoExplore Hanzo OS

Open sourceFully observableRun anywhereTechstars ’17

Replace the core. Integrate the rest.

Modern companies run on more than a hundred applications, and several hundred at enterprise scale. Most platforms add one more. Hanzo removes the need for whole categories of them: the applications a company shares all sit on the same company context, and the specialized systems that remain are reached through the integration layer.

Shipped first-party, each with a page you can open.
Communication
Projects and tasks
AI and agents
Software development
Product analytics
Observability
Infrastructure
Identity and security
Data services
Billing and commerce
What they share, and the reason the seams are not yours to maintain.
One identityOne project modelOne policy planeOne data contextOne operational traceOne bill
See what connects

The system

One OS. Every layer of the company.

Most companies assemble AI, communication, project management, analytics, infrastructure, identity, security and billing from separate vendors, then pay to make them behave as one thing. Hanzo was designed as one system — four layers, and two rails that cut across all of them.

Learn & adapt
Insights, traces, evaluations, experiments and outcomes — a rail across every layer, not a layer of its own.
Trust & govern
Identity for every human and every agent, permissions, secrets, boundaries, approvals, audit and post-quantum.

One company context.

One intelligence layer.

One agent runtime.

One policy plane.

One operational record.

Where people and AI work together.

Agents work inside the company, not outside it in another chat window.

One workspace
Projects, tasks, issues, channels, messages and documents belong to a workspace. People and agents read and write the same records, so work never has to be copied out of a tool and into a chat to reach whoever does it next.
An agent is a member
A bot joins as an ordinary member with a role, resolved from a Hanzo IAM service account or the agent registry. It can hold a task, answer in a channel, and be given or refused access the same way a person is.
Work carries a state
A task moves through backlog, in progress, review and done, with a priority, an assignee and a reporter. What an agent did is a record you can sort and filter, not a message somebody has to read to find out.
Authorship outlives the agent
An agent taken off a workspace is deactivated rather than deleted, so every message and document it wrote keeps its author. The history still reads correctly a year later.
See Hanzo Team
Enso · the intelligence layer

Intelligence at the center.

Enso reads the objective, understands the context, and puts the right intelligence behind the work. Fast when it can be. Powerful when it needs to be. Governed by your company’s policies, through one API across the system.

Try Enso View measured results
FlashProUltraone OpenAI- and Anthropic-compatible API

Agent runtime

Agents with hands.

Hanzo agents do more than answer. They use tools, operate software, browse, write code, work with files, run workflows and hand off to each other — inside isolated execution environments. Tabs puts every one of those shells in front of you.

Explore TabsExplore agents

Hanzo Cloud

AI-native all the way down.

500+ models behind one endpoint. Compute, GPUs, functions, SQL, vector, KV, document, object and analytical data, messaging, durable tasks, networking, identity, secrets, deployment, metering and billing — built together, reached the same way.

Kubernetes
The substrate everything runs on — your cluster or ours, same operator, same resources.
Visor
Schedules across clouds rather than inside one. Tier-1 and tier-2 providers are first-class, not adapters.
Firecracker
microVM isolation with a hardware boundary, for work that must not share a kernel.
gVisor
A user-space kernel where the boundary matters more than the last microsecond of start-up.
Functions
Short-lived workloads that scale to zero and bill by what they used.
Sandboxes
Where an agent is given hands — a filesystem, a terminal, a browser and a network it cannot leave.

Run Hanzo OS on Hanzo Cloud or your own infrastructure.

Explore Hanzo CloudHow Visor schedules it

Hanzo Insights

A company that learns from every outcome.

See what customers did. See what agents did. See which decisions produced results — then improve the memory, the workflow, the routing or the policy that produced them.

  1. Sense
    Events, customer behaviour, company data, product usage.
  2. Understand
    Memory, context, retrieval, analysis.
  3. Decide
    Enso, routing, planning, policy.
  4. Act
    Agents, tools, Tabs, sandboxes, workflows.
  5. Measure
    Traces, cost, evaluations, experiments, business outcomes.
  6. Adapt→ back to Sense
    Memory, prompts, workflows, routing, policy, allocation.
Explore Insights

See the whole company think.

One trace carries the whole path: the request that started the work, the services, queues and databases it crossed, the model calls it made, the tool calls under them, and what each one cost. It is one store, so a span and the log line it produced are one query rather than two tabs and a copied timestamp. Every panel scopes to your org from the token.

One trace, end to end
Follow one request across services, queues and databases, then open the span that spent the time. Child spans nest under their parent, so a model call and the tool calls beneath it read as one tree.
One store, one query
OpenTelemetry lands in a single column store and comes back as APM, logs, traces, metrics and exceptions. A span carries the id its log lines were written with, so the pivot is a click.
Every model call
Each generation records its model, its provider, and its prompt, completion and total tokens. A tool call is an observation of its own type, under the call that made it.
Every cost
Cost is stamped on the span as it arrives, from per-model pricing rules you edit — input, output, cache read and cache write. It totals per call, per trace, per conversation and per end user.
Sessions
A conversation is every trace and observation that shares a session id, carrying its own token and cost total. The same rollup lands per end user.
Scores and review
Eval scores and human feedback attach to a whole trace or to a single observation, numeric or categorical. Human notes carry a queue and a status, so a judgement is recorded where the work is.
Infrastructure underneath
Hosts, processes, pods, volumes, nodes, namespaces and clusters, beside the deployments, jobs and stateful sets running on them — in the same store as the traces they served.
Dashboards and alerts
An alert rule is a saved query, so what fires at three in the morning is something you already looked at. It reaches PagerDuty, Opsgenie, Slack, Teams, email or a webhook.
Explore Hanzo O11y

Every agent has an identity.

Every action is authorized. Every secret is scoped. Every execution is isolated. Every approval is attributable. Every change is auditable.

One service issues every credential
Hanzo IAM is the only thing that holds a password or runs a login, and nothing else in the stack has its own. An application hands you to IAM and gets back a code. A client with no registered secret — anything running in a browser or on a laptop — must present a PKCE challenge, and only S256 is accepted.
A token is trusted for what it cryptographically is
Bearer tokens verify against a closed list of signing algorithms: RSA, the NIST curves, and ML-DSA-65. HMAC and alg:none are absent from it. A refresh token is single-use, and presenting a spent one deletes its whole family rather than the one token.
Authorization is a grant at a place
A resource has a path: organization, then workspace, then project. One check asks whether some grant the caller holds covers the path and admits the verb. A grant is matched exactly rather than by prefix, so acme/prod does not cover acme/production, and the decision is a function call rather than a service that can be down.
A key is derived, not stored
Each database takes its key from one master through HKDF-SHA256, salted with the organization that owns it, so no two share a key and a file reopens after a restart with nothing kept beside it. Every secret gets a fresh key sealed under the master with AES-256-GCM and bound to its own path, so a ciphertext cannot be moved to another tenant. A master of the wrong length is an error, not a quiet fall back to no key.
One organization, one file
A tenancy boundary that only exists in a WHERE clause is one bug from being nothing. On Hanzo Base an organization’s data is its own database file, opened under its own derived key, so a query cannot reach across two — there is no second file open to reach into.
Execution runs behind a boundary it cannot lose
Every sandbox runs inside an isolation boundary, and the default is gVisor. A node that cannot provide the boundary that was asked for is an error rather than a substitution — nothing hands back a weaker runtime than the one requested — and the plain container runtime is reachable only for an organization named on an allowlist. Every Linux capability is dropped, and eight are handed back.
The edge deletes what a caller claims to be
Headers naming an organization, a user, an email or a role are stripped at the gateway before a handler reads one, and identity is written back only from a verified token. In transit it is TLS 1.3, and the edge offers hybrid ML-KEM-768 key exchange. Inside the cluster, services reach each other over a binary protocol rather than the public internet.
One row per request
The trail records the organization and the user who acted, the address they came from, the method and the URI they called, the action, the status the server returned, and the time. Rows recording a consent answer, or a credential issued or revoked, are reserved: the API refuses to create, correct or delete one. Evidence the subject of the evidence can write is not evidence.

Identity, authorization, KMS, zero trust, private networking, policy and audit are part of the operating system, not integrations added afterward. Built for regulated, private and sovereign environments.

Read the mechanisms
Check them yourself

One interface

One interface for people and agents.

Every capability is reachable the same way, whoever is asking.

AI agentHanzo TeamConsoleCLISDKAPIMCPInfrastructure as code

Same actions.

Same identity.

Same policies.

Same audit trail.

Everything else an app needs

Databases, storage, identity, secrets, agents, payments. Each one is a route on api.hanzo.ai/v1 — same key, same base URL. Open source, metered, and settled on-chain.

AI
Models · Providers · Agents · Inference · Embeddings · Playground · Prompts
Compute
GPUs · Machines · Functions · Edge · Applications · Clusters · Kubernetes · Tasks
Data
Vector · SQL · KV · Object Storage · Datastore · Base · DocDB · Memory
Network
Gateway · Nodes · VPC · DNS · Load Balancer · Service Mesh
Security
IAM · Authz · KMS · Secrets · Audit · Zero Trust
Dev
CLI · SDKs · API · Integrations · IDE · Desktop · API Keys
Infrastructure
Projects · Environments · Builds · Registry · Releases · Pipelines
Observe
Overview · Evals · Logs · Metrics · Traces · AI Metrics · Dashboards · Alerts · Billing · Plans & Pricing · Datasets · Experiments · Sessions · Scores · Score Configs · Annotation Queues
Web3
Settlement · Wallets · Referrals · Tokens · Networks · Indexer · Oracles
Apps
Chat · Bot · Marketplace · Search · Crawl · Studio · Console

Run it on your own machine

Most AI platforms hand you an API key. We hand you the platform. One binary, and hanzo up brings up the cloud itself — control plane, identity, secrets, gateway, storage, pub/sub — on your own hardware.

It is the same code we run in our own cloud, under an open licence. Point a client at localhost or at api.hanzo.ai — same API, nothing else changes.

curl -fsSL https://hanzo.sh | shInstall options

Desktop

The whole stack as an app. The agent node ships inside the bundle as a sidecar; inference comes from the engine, running locally or on your LAN.

DownloadSource

CLI

One static Rust binary — a coding agent, the MCP server, and every product of the cloud from your terminal. No runtime, no daemon.

InstallSource

Engine

The inference engine that actually runs the model. LLM and embedding serving in Rust, built for local hardware first.

DocsSource

ML

The compute core underneath: multi-backend tensors across CPU, CUDA, Metal, ROCm and Vulkan, with quantization built in.

DocsSource

The research is public

The methods, the measurements, and the results that did not work.

Cloud Economics of a Sovereign OSS StackJun 2026
A Regulated Capital-Markets Migration Case Study
One Native Stack for Private, Continuously-Learning AIJun 2026
Inference at the Bandwidth Wall, On-Device QLoRA, One Engine
Native ROCm Inference on a Consumer RDNA3.5 APUJun 2026
Reaching llama.cpp Decode Parity via a Unified 1-bit-to-Full Quant Core
Edge LLM Inference Across Commodity AcceleratorsJun 2026
Measured AMD, NVIDIA, and Apple Performance
Native Training in the Hanzo EngineJun 2026
One Engine, One Quant Format, One Backend
Continuously-Learning Private AIJun 2026
Self-Improvement and Privacy in the Hanzo Native Stack
Economical Serving and the hanzo.ai PlatformJun 2026
A Bandwidth-Optimal Unified Train-and-Serve Runtime
The Unified Sovereign-Tenant CloudJul 2026
Linear Shared-Nothing Scaling via Per-Tenant SQLite and In-Process Composition
Refutation-Driven Performance EngineeringJul 2026
An Empirical Study of a Multi-Week GPU Kernel Campaign
Evolutionary Schedule Search in a One-Source Kernel DSLJul 2026
And the Continual-Improvement Loop It Anchors
Hanzo RouterJul 2026
Memory-Aware, Local-First LLM Routing with a Learned SLO-Constrained Policy
Hanzo Network WhitepaperOct 2024
L1 Blockchain for Decentralized AI Compute
Read the papers

Companies have been running on this since 2014.

Hanzo OS started as one storefront in September 2014 and has been the system underneath our work ever since. More than a hundred venture-funded companies have built on it with us, through the Sensei Method — our engineers sitting with their team and working the problem alongside them. Among them, exits and multi-billion-dollar outcomes. These are the measured results after adoption.

TrillerFestThe largest virtual music festival held to that point — platform, production and ticketing.
Over 5 million viewers100+ artists across 3 days
Damon MotorcyclesCommerce and campaign stack for an electric motorcycle launch.
500x return on marketing investment230% more qualified leads45% lower acquisition cost
Personas SocialFeed, ranking and moderation for a social platform.
400% more content engagement3x feed relevance60% lower moderation cost
UnikrnToken launch, end to end, under real load on launch day.
120,000 ETH raised300% more platform engagement15+ jurisdictions
BellabeatThe platform behind a women’s health wearable.
67% more user engagement45% better retention
Cover BuildConfigurator and ordering path for factory-built housing.
300% more qualified leads40% shorter sales cycle
Myle TapLaunch of a wearable, from crowdfunding through fulfilment.
200% of the crowdfunding goal
Casper LabsTechnical architecture, launch and validator operation for an enterprise-facing chain.
Read the case studies

The model, the app, and the cloud it runs on

One login covers all three, and the API is the same one you build against locally.

01

Start with Enso

Hanzo’s frontier model. It reads each request and puts the right models on it, through one OpenAI-compatible API. Three presets: Flash, Pro, and Ultra.

Explore Enso
02

Build the app around it

Describe what you want in Hanzo App and it writes the UI, the schema, the auth, and the API. Keep chatting to change it, or edit the code directly.

Open hanzo.app
03

Run it on Hanzo Cloud

Zen open models, Base backends, IAM, KMS, and vector search, behind one API. Pay for what you use, or run the same code on your own hardware.

Explore Cloud