AI Dispatch with Claude as Quality Floor

CLAUDE-VERIFIED FLAT PRICING MCP-NATIVE NO SUBSCRIPTION

Cheap models handle bulk dispatch. Claude verifies the output.

Your agent gets Haiku/Mini/Flash speed at their price, with Claude-grade quality assurance. If the cheap path falls below the quality floor, we automatically escalate.

The dispatch flow

1. Your agent sends an intent.
2. Decomposer routes to the cheapest viable model (Haiku 4.5, GPT-4o-mini, or Gemini Flash).
3. Claude verifies the output against the task class quality floor.
4. If verification passes → return result, charge the flat task price.
5. If verification fails → escalate to a stronger model (Claude Sonnet/Opus). Same flat price.
6. Result + quality score + cost breakdown in response.

You see one price. Claude sits at the heart of the workflow as the truth oracle.

Pricing (live, transparent)

Task	Flat price	Quality floor	Claude verifies?	Production runs
detect_language	€0.005	0.99	Yes	231
classify_sentiment	€0.005	0.95	Yes	339
extract_emails	€0.005	0.95	Yes	187
extract_phones	€0.005	0.90	Yes	32
extract_keywords	€0.007	0.90	Yes	9
summarize_to_one_sentence	€0.007	0.85	Yes	20
translate_short	€0.007	0.85	Yes	4
ship_protocol_feature	€0.050	0.90	Yes — compound	49
Live manifest: /capabilities/dispatch/manifest

Why Claude as the quality floor

We measured. In our published Cross-Provider Quality Dataset (Q2 2026):

Claude Haiku 4.5 leads on latency (823ms mean across all task classes).
Claude scores 100% on detect_language and extract_emails in our production traces.
Cross-provider accuracy is tightly clustered — routing logic, not model selection, determines outcomes.

So we let the cheapest viable model do the work, and give Claude the verification authority. Claude approves the result or asks for re-runs. You pay one flat price. We absorb the cost of escalation when needed.

3 steps to dispatch

1. Register agent (free, no card)

curl -X POST https://astranl.com/api/agents/register \
  -H 'Content-Type: application/json' \
  -d '{"name":"my-agent","contact":"you@example.com"}'
# returns api_key starting with ask_*

2. Top up wallet (€1 = ~200 calls)

curl -X POST https://astranl.com/api/agents/wallet/topup \
  -H 'X-Agent-Key: ask_YOUR_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"amount_cents":100}'

3. Dispatch

curl -X POST https://astranl.com/capabilities/dispatch \
  -H 'X-Agent-Key: ask_YOUR_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"intent":"detect language: bonjour mon ami","quality_floor":0.95}'

Response includes:

{
  "result": "french",
  "task_class": "detect_language",
  "quality": 1.000,
  "verified_by": "claude-haiku-4-5",
  "billing": {
    "price_eur": 0.005,
    "floor_met": true,
    "escalated": false
  }
}

What we publish

Cross-Provider Quality Dataset Q2 2026 — 813 ground-truth runs, CC-BY 4.0
Live manifest — current pricing, task classes, quality floors
llms.txt — AI-readable summary
MCP Registry listing — verified server
Glama.ai listing — community directory

MCP integration

AstraNL is a verified MCP server. Add to Claude Desktop, Cursor, or any MCP-aware client:

{
  "mcpServers": {
    "astranl": {
      "transport": {
        "type": "streamable_http",
        "url": "https://astranl.com/mcp/streamable/"
      }
    }
  }
}

SSE legacy endpoint also available at https://astranl.com/mcp/sse.

Get started

AstraNL ZZP · KvK 88449335 · BTW NL004604224B69 · GDPR-compliant · Stripe billing · prepaid wallet · no subscriptions · cancel anytime by stopping calls.
Quality benchmarks published openly under CC-BY 4.0.