Agent-native access · v0.2 · same data, zero scraping

The registry, inside your agent

Agents are starting to shop for compute. Before one recommends or buys capacity, it should be able to ask three things: what meets my requirement, will it actually show up, and is this contract any good. Compute Assay answers all three as MCP tools, from this week's registry and the hash-chained delivery ledger, with the citation attached.

Connect

claude mcp add --transport http compute-assay https://computeassay.com/api/mcp

Or any MCP client (Cursor, VS Code, custom agents): Streamable HTTP endpoint at https://computeassay.com/api/mcp. No auth, no key. Same CC BY 4.0 terms as the open data: attribute Compute Assay. Health check: /api/mcp?health=1.

Agent tools (new in v0.2)

Tool	What it answers	Example ask
`find_capacity`	Offerings that meet a requirement (accelerator, contiguous GPU count, price ceiling, fabric, minimum grade, on-demand or term), cheapest first, each with what the seller does not publish	"64× H100 on InfiniBand under $4 an hour"
`delivery_risk`	For an announced AI site, or all of them: grade, probability the next phase is energized on time and within six months, power status, latest delay signals, and the ledger hash of the forecast	"Will Stargate Abilene's last buildings be ready this year?"
`grade_contract`	One contract marked the way the Commitment Book does: price against a competitive replacement rate, indicative value over the remaining term, delivery risk, acceptance, transferability, alerts. Nothing is stored.	"512 B200s at $6.40 for two years from Barber Lake — good deal?"

Registry tools

Tool	What it answers
`list_accelerators`	What's tracked, seller counts, price ranges
`gpu_prices`	Every listing for a GPU: price, grade, fabric, terms, source
`fair_price_band`	The negotiation anchor: Grade A/B walk-up band + cheapest committed rate
`provider_record`	One provider: offerings, published SLA/terms, what they don't publish, 90-day incident history
`model_fit`	Cheapest GPUs that fit N-billion params at a precision
`cost_per_million_tokens`	Rented $/M output tokens from sourced benchmarks, assumptions stated

What an agent should know

Grades describe disclosure, not measurement, until a fleet is Verified by the harness. Every response says so.
Posted prices are not availability. find_capacity flags listings whose largest contiguous allocation is unpublished. Confirm with the seller before committing.
Delivery forecasts resolve in public. Each carries the hash of its ledger entry, so an agent can cite exactly which forecast it relied on.
Compute Assay never brokers or executes. The tools inform a decision; they don't place orders. That's what keeps the grades neutral.

Every response carries a citation field and snapshot date. If you build on it, keep both. Methodology · delivery risk · open data.