Settlement for AI work.
The open settlement and coordination fabric underneath AI work, beneath the inference and compute providers, model builders and agent platforms of the machine economy. Each call is routed, authorised, settled and returned with a receipt anyone can check.
- Signed by hardware
- Every answer a registered machine serves from a catalogued model carries a receipt signed under that machine's hardware key.
- Billed by what was sent
- A call is billed from the provider's signed receipt, capped by what was actually sent and received.
- Revoked everywhere at once
- Revoke an agent and nothing signed for it after the revocation is paid, on any operator's channel.
- Open source
- The node, the SDKs and the protocol are open source under Apache 2.0.
Agents now pay for work.
Software that acts on someone’s behalf now calls models, buys data and hires other software, often many times a second.
Invoices and card payments were built for people buying from companies. They do not say which agent spent, under whose authority, or what the money bought.
- Participants
- Humans
- Agents
- Machines
- Services
- Rails
- Settlement
- Consensus
- Hardware
- TPM 2.0
- Secure Enclave
- Passkeys
- TEEs
- GPUs
Operators have the machines. Tenzro brings the work.
- Routed
To the provider that fits
Calls go to the provider that fits by price, latency and reputation, and move on within seconds if one fails. - Settled
Paid per call
Each call is paid against a receipt the serving machine signs, with billing capped by what was actually sent and received. When a job spans several operators, each is paid its share. - Recorded
One record per payment
Every payment is one record that binds the request, the quote and the price. Refunds and revenue splits settle on the same record.
Every agent answers to someone.
Agents now act without authorisation, and leaked machine credentials are this year’s main attack path.
Every agent on Tenzro answers to a person and holds no keys. Its authority is a set of terms that person grants with a passkey, enforced by the protocol rather than by prompts, and every payment and tool call it makes is recorded against those terms.
Terms, not keys
An agent's authority is its terms: what it may pay, in which assets, up to which per-payment and daily limits, granted by a person's passkey or a machine's hardware key. High-frequency work over escrow-backed payment channels counts against the same limits.Revocation everywhere at once
Revoke an agent and nothing signed for it after the revocation is paid, on any operator's channel.A signed record for every tool call
Every tool call an agent makes carries a signed record: which agent acted, under whose authority, what was allowed or refused, and any human approval.Evidence an auditor verifies offline
An agent's payments and tool-call records export as an evidence pack an auditor verifies offline, without trusting the operator. These are records built for regulatory oversight of AI, such as the EU AI Act's logging and human-oversight duties.
Every class of machine has work to do.
- Edge
Small and embedded devices
Single-board computers, Jetson-class modules, mini PCs, NAS boxes and gateways. They hold identity, make payments, run agents and storage, and pass inference on to GPU providers. - Provider
CPU servers and workstations
Every role a CPU can serve, including the models that fit it. - Provider
From one GPU to a data centre
Inference, compute rental and confidential computing, with models too large for one machine served in pipeline shards across your own machines, and mixture-of-experts models across holders on the network. - Validator
Servers with a TPM
Order transactions and certify checkpoints with keys held in the hardware, against a bonded stake.
One network for everyone who buys AI work.
Developers building AI applications
Leading open-source models behind OpenAI-compatible APIs, priced by an open market instead of one provider's rate card.Learn moreAgents and machines
Hardware-rooted identities, spending limits and metered payments, so software can buy inference and compute on its own.Learn morePrivate and sensitive workloads
Run inference and data processing inside attested hardware, with evidence you can check yourself.Learn moreRegulated industries
Choose where models, data and compute live, and see who served every request, under which policy, certified by whom.Learn moreModel builders and researchers
Rent capacity, and train across distributed clusters with applications built on the network.Learn moreEvery industry that buys AI
One endpoint for inference and compute from independent providers, without a contract with each of them.Learn more
Post-quantum signatures, hardware-rooted keys, agents bounded by their owner.
Built to be trusted by machines.
- 01Tenzro DAG Consensus (TDC)Our post-quantum consensus protocol. Every validator proposes in parallel each round, so no single leader limits the network, and its records stay verifiable for decades.
- 02Keys that live in hardwareEvery key is rooted in a TPM, a Secure Enclave or a passkey. No private key is ever stored in the clear, and there are no seed phrases.
- 03Trust and provenanceConfidential compute on any cloud or on premises, with attestation bound to each job's receipt, sealed model weights, and certification issued by anyone. You choose which issuers you trust and check the rest yourself.
- 04The Machine EconomyA compute price index signed by validators from what their routers observe, SLA-attested rentals and metered, per-use settlement, paid in TNZO or stablecoins, with circuit breakers and formally proven money arithmetic.
- 05Inference for any modelPrefix and state reuse for every architecture class, behind OpenAI-compatible APIs for chat, responses, embeddings and audio, with image and video generation on the same node's API.
- 06Closed by defaultOpen, owner and admin methods are separate. Anything that moves money or changes state needs a signature from the account that pays.
Open source under Apache 2.0.
An economy for distributed compute and decentralized AI, open to machines of every size, from home nodes to data centres.
Run a nodeCompute on a machine you control.
Install one binary. Its keys sit in the machine's TPM or Secure Enclave, and a private model is never announced or served off the node. Serve a model, or just follow the network and pay.
Provide computeBe paid for the work it does.
Calls and rented hours settle with receipts anyone can verify. Pricing is yours. What a provider earns follows the paid work it serves; nothing is paid for holding hardware, being listed or signing up.
Home labsPut the home lab to work.
Start with hardware you already have, from a Pi or a home computer to a repurposed server or workstation; no new facility. A node is software on hardware already in place: download, verify, run tenzro setup. Small devices follow the chain and pay; serving roles need hardware that can serve, and what a provider earns follows paid demand.
BusinessesCompute you can account for.
Call independent providers through OpenAI-compatible APIs and pay per use, with a signed receipt for each call. Choose by price, latency and standing, and pin a session to a jurisdiction. A business with machines of its own can add them: a node is software on hardware already in place.
Enterprise and operatorsBring the stack you already run.
Operators and AI factories are supply and customers of the network. The network adds capacity from existing machines and does not replace data centres: one identity and payment rail across all tiers, including them. A small data centre or server room can join with a node, or with a connector in front of the stack it already runs, and publish a signed operator policy. Payment is per completed unit of work. Partner platforms can run the node for you without holding your keys.
Distributed computeAdd capacity from machines that already exist, alongside data centres.
Work is routed across independent providers, from a home node to a data-centre rack, and fails over to another provider if one drops. A node is software on hardware already installed: no new building, facility lease or construction permit, and no new grid connection. Operators still follow their own local rules. Larger models are designed to be split across expert holders, each settled for its own work. One identity and payment rail runs across every tier, including data centres.
PlacementChoose where your compute runs.
Providers declare a country and a jurisdiction, and a session can be pinned to one. Keys stay in hardware and receipts are verifiable. A declaration is the operator's, not an audit, and not a statement of legal compliance.