Early access — Compute markets intelligence

GPU-hour evidence inside your coding agent.

Read current GPU-hour evidence, inference economics, power readiness, and source-linked signals over MCP.

Source citedThreshold clearedDesk reviewedHuman approved

Get early access

See live signals

Why MCP

The terminal is moving inside your AI assistant.

Your agent inspects a repository locally, reduces the configuration to a fixed set of normalized facts, and calls GreenCIO with those facts — never with source text. Compute market state, inference economics, power readiness, and source-linked evidence arrive in the same session, each with sources attached.

Two profiles, one server

Know which profile you are connecting.

One binary, selected by --profile. Early access is the coding profile: read-only, key-scoped, and what every example on this page uses.

--profile coding

Customer surface

Thirteen tools, every one annotated read-only.
No write tool is registered, so no agent action can mutate GreenCIO state.
Repository access is bounded by an allowlist the server enforces, not by convention.
Tenant workspace actions are out of scope and require a separate authenticated profile.

--profile analyst

Internal research surface

Twelve tools built for GreenCIO desk research, not for distribution.
Includes a URL-ingestion write tool that adds to the intelligence feed.
Remains the default profile when no flag is passed.
Not part of the early-access offer.

For analysts and capital allocators

Ask in plain English. Defend the answer with source trails.

You

What is the H100 US-East spot settlement right now?

→ get_compute_market_snapshot

greencio MCP

{
  "settlement": {
    "indexName": "GCI-H100-USE-SPOT",
    "sku": "H100",
    "region": "US_EAST",
    "tenor": "SPOT",
    "asOf": "<observation time>",
    "value": "<current settlement>",
    "bandLow": "<band low>",
    "bandHigh": "<band high>",
    "provisional": false
  },
  "methodologyStatus": "persisted",
  "methodologyVersion": "<methodology version>",
  "generatedAt": "<generated at>"
}

You

Find source-linked Meta datacenter evidence from this month.

→ search_ai_infra_evidence

greencio MCP

{
  "signals": [
    {
      "title": "<publisher headline>",
      "sourceUrl": "https://publisher.example/article",
      "sourceName": "<publisher>",
      "publishedAt": "<published at>",
      "impactLevel": "high"
    }
  ],
  "totalMatched": "<count>",
  "dataMode": "gcs",
  "warnings": ["Search results are source-linked evidence leads. They do not prove causation or deployed capacity."]
}

You

Explain the H100 move over the last seven days.

→ explain_compute_price_move

greencio MCP

{
  "indexName": "GCI-H100-USE-SPOT",
  "window": "7d",
  "move": {
    "firstValue": "<first settlement>",
    "lastValue": "<last settlement>",
    "absoluteChange": "<change>",
    "percentChange": "<change %>",
    "firstObservedAt": "<observation time>",
    "lastObservedAt": "<observation time>"
  },
  "relatedSignals": ["<source-linked evidence lead>"],
  "methodologyStatus": "persisted",
  "dataMode": "gcs",
  "warnings": ["Related signals are temporally or topically associated evidence leads. They do not establish causation."]
}

Field names are exact; every value is a placeholder resolved by the connected host. This page does not publish current settlements. Each response carries the source and method context the assistant needs to say where a number came from.

For engineering buyers and AI infra teams

Thirteen decision tools. One shared evidence contract.

Coding profile

analyze_ai_infra_repositorypro keyReview normalized facts from local infrastructure files.
get_compute_market_snapshotpro keyGet a current Compute Index market snapshot.
get_compute_price_historypro keyMeasure a market move across historical settlements.
compare_compute_marketspro keyCompare two current Compute Index markets.
get_compute_methodologypro keyInspect price method and qualification context.
list_model_token_pricespro keyList model input and output token-price benchmarks.
estimate_inference_economicspro keyCalculate unit economics from submitted assumptions.
calculate_lcoepro keyCalculate levelized cost of energy from submitted assumptions.
screen_time_to_powerpro keyRun a heuristic power-readiness screen.
explain_compute_price_movepro keyPair a measured price move with non-causal evidence context.
search_ai_infra_evidenceno keyFind source-linked infrastructure evidence leads.
estimate_datacenter_paybackno keyEstimate simple project payback from submitted assumptions.
analyze_gpu_hour_optionno keyAnalyze a GPU-hour option contract from submitted assumptions.

Ten tools reach the authenticated Data API and need a pro-tier key. Evidence search reads the public feed; the payback and option models are pure functions of submitted assumptions. A missing key is a call-time UPSTREAM_AUTH on the tools that need it, not a startup failure — the rest keep working.

Architecture

Two transports. stdio for a local subprocess, or a self-hosted Streamable HTTP listener for a shared internal deployment. There is no GreenCIO-hosted remote endpoint yet.
Bounded upstream calls. A fifteen-second request timeout, one retry on an upstream 5xx, and a per-response character budget so a tool result cannot flood the model context.
Structured errors. Every failure returns a machine-readable code and, where useful, the upstream HTTP status and a remediation hint — never a partial or invented result.
Sources included. Feed responses expose dataMode; Compute Index responses expose methodologyStatus and methodologyVersion.
No invented data. Tools return structured errors or explicit warnings instead of fabricated forward curves or settlements.
Read-only by construction. All thirteen coding tools carry the read-only annotation and none of them writes.

Error codes

Code	Raised when
`BAD_INPUT`	Arguments failed the tool's own schema before any network call.
`UPSTREAM_AUTH`	The Data API rejected the key, or no key is configured.
`UPSTREAM_NOT_FOUND`	No settlement, spread, or resource matches the request.
`UPSTREAM_RATE_LIMIT`	The Data API burst or monthly quota was exceeded.
`UPSTREAM_5XX`	The Data API failed after one retry.
`UPSTREAM_TIMEOUT`	The upstream call exceeded the request timeout.
`SYNTHETIC_BLOCKED`	Synthetic Compute Index data was refused under STRICT_NO_SYNTHETIC.
`LOCAL_RATE_LIMIT`	A local per-process guard on the analyst write tool tripped.
`NOT_IMPLEMENTED`	The requested capability is declared but not shipped.
`INTERNAL`	An unexpected server-side failure.

Repository data boundary

Seven fields leave your machine. The schema enforces it.

A repository review sends a relative path, an optional line number, a source kind, and at least one of these normalized facts per finding:

provider
region
gpu_type
gpu_count
replicas
utilization_rate
model

The shape is enforced on both sides of the wire. An unrecognized property on a scan or a finding is rejected by name on your machine, before any request is built; a stray property outside the scan object is dropped at the tool boundary and never serialized. The Data API then validates the same schema again on arrival and answers 422. There is no field an extra property can ride in on.

Never sent: source text, code snippets, environment values, credentials, secrets, personal data, or business logic. Paths must be repository-relative — an absolute path or one that traverses out of the repository is refused by the schema, because a home directory or share name discloses a person and a project we never asked about. Every review response restates the boundary as an explicit provenance assertion, so an audit reads the claim off the record rather than off this page.

Scanning and redaction happen on your side. The CLI is the reference implementation of that boundary and can run with network access disabled to produce a scan you inspect before anything is sent.

What you don't get from a UI alone

Defensibility for the answer the chat agent gives.

Source trails

Every signal carries source URL, source name, and discovery timestamp. The chat agent can quote evidence back into a memo without losing the citation.

Methodology versioning

Every Compute Index result identifies the method revision it was settled under. The get_compute_methodology tool returns the active method, its rulebook hash, and its effective dates on demand.

No silent demo data

Synthetic upstream data is labeled methodologyStatus="synthetic" and carries a disclosure warning — or is refused outright with SYNTHETIC_BLOCKED when STRICT_NO_SYNTHETIC is set.

Connect it

Codex, Claude, Cursor, or any compatible MCP client.

Local subprocess (stdio)

Build the package once from a checkout — npm ci then npm run build inside mcp-server/ — then point your client at the compiled entry point.

{
  "mcpServers": {
    "greencio": {
      "command": "node",
      "args": ["/absolute/path/to/greencio/mcp-server/dist/index.js", "--profile", "coding"],
      "env": {
        "GREENCIO_BASE_URL": "https://www.greencio.com",
        "GREENCIO_DATA_API_KEY": "gc_live_...",
        "STRICT_NO_SYNTHETIC": "true"
      }
    }
  }
}

Shared internal deployment (Streamable HTTP)

The listener binds to loopback by default. Binding to any other host requires a bearer token, and the process refuses to start without one.

# Self-hosted Streamable HTTP listener
GREENCIO_DATA_API_KEY=gc_live_... \
GREENCIO_MCP_BEARER_TOKEN=<long random token> \
node mcp-server/dist/index.js --profile coding --transport http

# MCP endpoint  http://127.0.0.1:8787/mcp
# Health probe  http://127.0.0.1:8787/health

Distribution status

Surface	What it is	Status
MCP coding profile	Thirteen read-only tools over stdio or self-hosted Streamable HTTP.	Implemented; npm publication pending
GreenCIO CLI	Local repository scan and redaction. The primary repository boundary — it is what decides what leaves your machine.	Implemented; npm publication pending
Portable agent skill	A client-agnostic skill file that teaches an agent when to reach for the CLI versus a specific MCP tool.	In repository
Codex plugin	The bundled install unit for Codex, wrapping the same skill and MCP configuration.	In repository; marketplace publication pending
Managed remote endpoint	A GreenCIO-hosted MCP endpoint with per-tenant authorization.	Not deployed
MCP Registry entry	Discovery under com.greencio/ai-infrastructure, pending namespace verification.	Not published

Server identity is @greencio/mcp v0.3.0, registered as com.greencio/ai-infrastructure. The pro-tier Data API key is minted by GreenCIO during reviewed early-access onboarding.

Early access

What to expect right now.

When is it live?

The repository implementation and local test suite are complete. npm publication, MCP Registry publication, managed remote deployment, and client compatibility proof remain release gates. Early-access cohorts run against a local checkout.

What do I need?

Node.js 18 or later, a local checkout, and a pro-tier GreenCIO Data API key for the ten tools that reach the Data API. Ask for the products you intend to call when the key is reviewed — entitlements are set on the key, not implied by its prefix.

Which AI tools does this work with?

The server implements standard stdio and Streamable HTTP MCP. Codex, Claude, and Cursor provide MCP configuration paths. Pi can use the portable skill and CLI; Pi MCP use depends on an installed MCP extension. Client-specific release proof is still required, so treat any specific client as unverified until we have tested it with you.

Is data sent anywhere new?

Repository tools send relative paths plus the seven allowlisted facts, and nothing else. Compute-market tools call the GreenCIO Data API with your key. The payback and option models call public GreenCIO calculators with only the assumptions you supply. No tool posts your repository anywhere.

Can we self-host it inside our network?

Yes — run the Streamable HTTP listener on your own infrastructure with a bearer token and your own TLS termination. Outbound calls still reach the GreenCIO Data API, so treat it as an egress-controlled service rather than an air-gapped one.

AI infrastructure evidence, inside your coding agent.

Join the early-access cohort. We'll set up your key and a supported local MCP client.

Get early access See live signals