What is live, and what comes next.

These capabilities run on Robinhood Chain today. Because every lease and settlement is recorded onchain, the record of what works is public rather than asserted.

Wallet-native access

An agent proves control of a wallet with a signed challenge and replay protection. No account, password, or human console stands between it and a machine.

Python and TypeScript SDKs

Price capacity, fund a lease, provision a machine, and open a session in a few lines. Published on PyPI and npm.

Model Context Protocol server

Any MCP client, Claude included, can list capacity, rent a GPU, and run a command through Prism. Listed in the official MCP registry.

Prism inside the agent frameworks people already use

An agent running Hermes gets three skills, rent a GPU, reproduce a CUDA run and read the receipts, from hermes plugins install prismnetwork-tech/hermes-plugin-prism. The published Python package is a version behind and carries the terminal backend without those skills, so the plugin install is the one to use. The Coinbase AgentKit action provider is published and installable, so an agent built on AgentKit, LangGraph, or the Vercel AI SDK reaches the same capacity with no custom code.

Spend caps an agent cannot cross

Every paid call states the most it can cost before the agent pays it, and the SDK keeps a running ledger against a session budget. An agent left to run unattended has a ceiling, and the receipt afterwards separates what was charged from what came back.

Prism Chain

A Layer 3 built for GPU compute and settled on Robinhood Chain, with its own explorer and bridge. ETH, USDG and PRISM move between the two chains through the canonical Arbitrum token bridge. Deposits arrive in about a minute; withdrawals finalize in about a day. The TypeScript SDK exports the chain and the bridge calls. Leases still settle on Robinhood Chain, and moving them is the next step below.

A stated reason behind every spend

An operator writes a policy: which actions an agent may fund and how sure it has to be. The agent's decision is checked before anything is quoted, and its hash is recorded with the escrow deposit, so the lease onchain points back to the reason it was funded. Both SDKs and the MCP server support it.

Onchain lease escrow and settlement

Funding, metered billing, and refunds execute through deployed contracts, bounded by a maximum deposit, a maximum duration, and a dispute window. Routine administration of those contracts sits behind an enforced 48-hour timelock, so a change to who signs settlements or where fees go is scheduled in public two days before it can execute.

Ending a lease early

A renter closes the session the moment the work is done and the meter stops there. Settlement charges the time the machine was actually open and returns the rest of the escrow. One recent lease held a workstation GPU for twenty-eight seconds, charged 0.006216 USDG and returned the remaining 0.126984 on the same receipt.

Verifiable receipts

Every finalized lease publishes a settlement receipt that anyone can confirm on a block explorer.

Concurrent multi-class GPU capacity

Supply is sourced from vetted providers across several NVIDIA classes, including L40S, RTX A6000, RTX 6000 Ada and RTX 5880 Ada, and matched to a lease on demand. Leases run alongside each other rather than one at a time. How many machines are on offer at any moment is a small number, usually one to three. Raising it is the first item below.

Managed and confidential inference

Text generation paid per call over x402, in USDG on Robinhood Chain or USDC on Base, with no account or API key. Two models on the open tier and nine on the confidential tier, which runs the model inside a hardware enclave with an attested GPU. The client checks the enclave's attestation before the prompt is sent, so the prompt is readable only inside the enclave.

Findable by an agent that already pays over x402

Pay-per-job GPU execution and the inference endpoint are both indexed in Coinbase's x402 Bazaar and on agentic.market. An agent that speaks the payment protocol can find the endpoint, read the price, and pay it without anyone integrating Prism first.

A pinned session key per lease

Clients pin the SSH host key of the machine they rent for the life of the lease. Where no honest key can be published, because a third-party cloud generated it, the lease says so and a client can refuse it, rather than being told a check happened that did not.

An availability commitment on every lease

A machine that stops answering is billed only to the moment it was last observed. The receipt names the lease interrupted and states the seconds held but not charged, publicly, on the same receipt anyone can verify.

An encrypted vault that outlives a lease

Credentials and small artifacts persist across leases, encrypted to the owning wallet. Each item names the weakest workspace class it may enter, and a release below that floor is refused rather than logged.

A stated trust class per offer

Every offer, quote, lease, and receipt carries what the supplier protects, derived by the network from evidence it can check rather than asserted by the host. An agent can require a minimum class instead of reading a disclaimer. All capacity live today is the weakest one, and the security page states what that does and does not cover.

The constraint on Prism right now is how much capacity is on offer at once, and what a lease costs to run. That is where the work is.

More machines on offer

The filter that decides which rented machines qualify asks for more memory than most of the market has. On 7 September, nine machines across the supplier marketplace passed it. Lowering the floor to 24 GB and accepting more card models takes that pool to about sixty. You will see it as ten or more distinct offers on the public offers endpoint and on this site, held for a week, with at least one settled receipt from a 24 GB card in the public feed. Confidential inference keeps its current hardware floor, because the enclave needs hardware consumer cards do not have.

A quote that resolves to a machine

A funded lease reaches a shell inside four minutes when a host is there to take it. The work now is confirming a rentable host at quote time, so a lease is only opened against capacity that has already been checked. Escrow is still funded before anything is provisioned, and that does not change.

A LangChain package

The LangChain integration is finished and its last dependency is now on PyPI. It ships as a standalone open source package, installable with pip. The name langchain-prismon PyPI belongs to an unrelated project, so Prism's package publishes under a different one and this page will carry it the day it is up. LangChain stopped taking integrations into its own project in March, so a standalone package is the only shape this can take.

Prism in the tool catalogs

A listing on skills.sh, a listing on ClawHub, and an entry in the community Hermes plugin index. One package reaches Hermes, Claude Code, Cursor, Codex, Copilot, Windsurf, Gemini, Cline and about twenty more clients through the same install path.

Twenty offers, held for two weeks

Twenty or more distinct online offers on the public offers endpoint, continuously, for fourteen days, with at least 95% of advertised capacity provisioning successfully when leased. Advertised capacity counts only if it is there when a lease arrives for it.

Leases settle on Prism Chain

Robinhood Chain fees rose ninefold in eleven days while the lease price stayed fixed, which makes short leases expensive to settle. The escrow and settlement contracts move to Prism Chain, funded with USDG bridged from Robinhood Chain, where a transaction costs a fraction of a cent. The target is under one cent of chain cost per settled lease, held for thirty days. Prism Chain posts its data to Robinhood Chain, so the public trail stays there.

Spend controls, written down

A public page stating the per-call and per-session caps an agent can set, what happens when one is hit, and how to export receipts for accounting.

Directional commitments that widen what the network can be trusted to run. Each ships only when it meets the same onchain, verifiable standard as the rest of Prism.

A second supplier

One broker supplies all of Prism's capacity today. The goal is offers sourced from two independent brokers and a published drill showing at least five offers still served with the first supplier switched off. The reason is continuity: capacity a customer can rent should not depend on one company staying available.

Receipts a stranger can check without us

A published verifier that takes a receipt, checks the settlement records on chain and the enclave attestation behind them, and fails on a mismatch. It states which parts of a receipt it verified and which parts the supplier reported, so the difference is visible to whoever is reading it.

Own hardware, once usage earns it

Independent nodes with exclusive GPU passthrough are the only route to the isolation that rented capacity cannot promise. They are also a fixed cost that does not care whether anything is leased, so they start when the network books 400 leased GPU hours a month for two consecutive months. Until then all capacity is brokered and this site says so.

Attested and confidential workspaces

The two classes above what rented workspaces can verify today: a launch measurement checked against vendor roots, then hardware-backed trusted execution with encrypted GPU memory, so a workload and its data stay private from the infrastructure provider. Confidential inference already runs this way. Extending it to a workspace a renter controls waits on the hardware above.

Durable workspaces

Workspace storage that survives across leases. The vault carries credentials and small artifacts today; a full working disk that a new lease can resume is not built.

More framework integrations

Six surfaces already reach Prism. Another wrapper does not change what a customer can do, so a seventh is not scheduled.

Stronger service-level remedies

The live commitment is not being charged for an interrupted machine. Performance guarantees, and remedies beyond withholding payment, are not built and are not scheduled.

Interruptible leases in 2026

Preemptible capacity is roughly a third of the price and it breaks the duration a lease commits to. It needs its own product class before it can be sold, and that is a 2027 conversation.

Verifiable over asserted

Settlement is recorded onchain, so a claim about what the network did is checkable rather than taken on faith.

Agent-native by default

A wallet in, a machine out. Nothing on the critical path requires a dashboard, a card, or a person.

Honest boundaries

The security page states plainly what the network does not yet protect. This roadmap holds to the same standard.

Track what has shipped in the live receipts and in the documentation.