Gvnr

Spend caps, rate coordination, idempotency, and a human-in-the-loop gate for AI agents — enforced before the call, not after the invoice. One MCP endpoint, settled via x402 (USDC on Base). No proxy, no self-hosting, no infrastructure to deploy.

Listed on the Official MCP Registry as dev.gvnr/gvnr.

The problem

Agents cost 10–12× more than estimated in production. System prompts, retry loops, and tool calls multiply fast — a runaway agent can generate a $47,000 bill in 11 days. The usual fix, self-hosting a gateway like LiteLLM, means running infrastructure most developers won't set up.

Gvnr is the hosted alternative: an external authority your agent checks before it spends. Your agent asks "am I clear to make this call?" and acts on the answer.

How it works

Gvnr is not an LLM proxy — your tokens never pass through it, and you pay your model provider directly. Gvnr governs the decision to spend, in two independent meters:

1. The governance-operation quota — what you buy from Gvnr. Your account holds a balance of governance operations (operations_remaining). One budget_clear burns one op. You top this up with USDC; it's decoupled from your LLM spend. get_balance reports it.

2. The spend envelope — a USD cap you set, per agent. set_envelope(agent_id, limit_usd, window) gives an agent a daily or per-session USD ceiling. Gvnr tracks estimated spend against it and denies once it's exceeded — this is the runaway guardrail. No dollars move; it's an accounting limit you control.

A budget_clear is approved only when both hold: the account has ops left in the quota, and the agent is under its envelope. The loop:

1. Your agent calls budget_clear(agent_id, model, estimated_tokens) before each LLM request.
2. Gvnr checks the op quota and the agent's envelope, and returns { approved: true, ... } or { approved: false, reason }.
3. If denied, your agent skips (or escalates via request_approval).
4. After the LLM responds, call reconcile(...) with the real token counts so the envelope tracks actual cost, not the estimate.

Quick start

1. Provision an account

curl -X POST https://gvnr.dev/v1/account
# { "api_key": "bg_...", "account_id": "...", "operations_remaining": 25 }

New accounts include 25 free trial ops — enough to run the full loop (set an envelope, budget_clear, reconcile) before you fund anything. Your api_key is the credential for every call; account_id is just an internal reference for support. Account creation is credential-only by design (agent-native) — email is optional, set separately via POST /v1/account/notification-email, and used only for human approval notices.

2. Top up your governance-op quota

Pay-as-you-go: 1,000 ops per $1, any amount (try the whole rail for $1), USDC on Base. Open the pay page, name your amount, and pass your API key:

https://gvnr.dev/pay?usd=1&api_key=bg_YOUR_KEY

Send USDC to the address shown and paste your tx hash — ops are credited proportional to the amount received, after on-chain verification. Programmatic clients can submit the hash directly:

curl -X POST \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"tx_hash":"0x..."}' \
  https://gvnr.dev/v1/account/topup-verify
# { "operations_remaining": 1000, "credited_ops": 1000, "credited_usd": 1 }

Or settle in one round-trip with an x402 client (Base MCP, AgentKit, x402-fetch) — name your own amount:

POST https://gvnr.dev/v1/account/topup?usd=5      # → 402 challenge → pay → 5,000 ops credited

3. Set a spend envelope for your agent

curl -X PUT \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","limit_usd":5,"window":"daily"}' \
  https://gvnr.dev/v1/budget/envelope
# { "success": true, "agent_id": "my-agent", "limit_usd": 5, "window": "daily" }

4. (optional) Set a rate envelope per (agent, provider, model)

curl -X PUT \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","provider":"anthropic","model":"claude-sonnet-4-6","requests_per_minute":30}' \
  https://gvnr.dev/v1/rate/envelope

5. Before each LLM request: budget_clear, then rate_check

curl -X POST \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","model":"claude-sonnet-4-6","estimated_tokens":2000}' \
  https://gvnr.dev/v1/budget/clear
# { "approved": true, "remaining_usd": 4.994, "operations_remaining": 18999 }

curl -X POST \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","provider":"anthropic","model":"claude-sonnet-4-6"}' \
  https://gvnr.dev/v1/rate/check
# { "allowed": true, "requests_remaining_this_minute": 29 }

6. (optional) Dedupe retries with idempotency_check

curl -X POST \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"key":"job-abc-123","ttl_seconds":3600}' \
  https://gvnr.dev/v1/idempotency/check
# First call:  { "is_first_call": true,  "ttl_remaining_seconds": 3600 }
# Replay:      { "is_first_call": false, "ttl_remaining_seconds": 3598 }

7. After the LLM responds, reconcile against actual usage

curl -X POST \
  -H "Authorization: Bearer bg_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","actual_input_tokens":1800,"actual_output_tokens":2400}' \
  https://gvnr.dev/v1/budget/reconcile
# { "ok": true, "drift_usd": 0.003, "remaining_usd": 4.991, "operations_remaining": 18999 }

reconcile adjusts the spend envelope by the drift between your estimate and actual cost (the op quota is untouched). You don't pass the model again — reconcile reuses the one from your prior budget_clear. Anthropic, OpenAI, and Gemini all return usage fields with real token counts — pass those in to keep the envelope honest.

Human-in-the-loop approvals

When an agent hits a denial or a sensitive action, pause for a human instead of failing:

# 1. Open a request — returns an approval_id and a mobile-friendly approval_url
curl -X POST -H "Authorization: Bearer bg_YOUR_KEY" -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","action_summary":"Spend $40 on a research run","ttl_seconds":3600}' \
  https://gvnr.dev/v1/approval/request
# { "approval_id": "...", "approval_url": "https://gvnr.dev/approve/...", "expires_at": ... }

# 2. The human opens approval_url and taps approve / deny (emailed if notification-email is set).

# 3. Your agent polls until the decision lands:
curl -H "Authorization: Bearer bg_YOUR_KEY" \
  https://gvnr.dev/v1/approval/check/APPROVAL_ID
# { "decision": "pending" }  →  "approved" | "denied" | "timeout"

The agent proceeds on approved, skips on denied, and handles timeout (no decision before expires_at) however it likes.

MCP setup

Add to Claude Desktop or any MCP-compatible client:

https://gvnr.dev/mcp?api_key=bg_YOUR_KEY

Claude Code

claude mcp add gvnr --transport http \
  "https://gvnr.dev/mcp?api_key=bg_YOUR_KEY"

MCP tools

REST API

All endpoints except POST /v1/account require Authorization: Bearer bg_YOUR_KEY.

TypeScript users: generate full types from the live OpenAPI spec — npx openapi-typescript@latest https://gvnr.dev/openapi.json -o types/gvnr.d.ts. See TYPESCRIPT.md for the integration pattern (typed fetch, x402 topups, discriminated response unions).

Account & top-up