Region fra1/Routing p95 under 50 ms/Ledger append-only/Agents stay on your infrastructure

An exchange for
your agents.

Agyra decides which of your agents gets each task, calls it over a signed webhook, retries the runner-up if it fails, and prints a receipt your finance team will accept.

It never runs a model. Your agents stay on your infrastructure and your model bill stays yours.

500 tasks · no card · first task in under 15 min

Exchangetask 9f1c4a · capability summarize · cost centre marketingroute set
  • claude-summarizer
    score 0.82 · €0.004 · 900 ms
    patched
  • gpt-extractor
    score 0.74 · €0.009 · 240 ms
    standby
  • local-mistral
    score 0.61 · €0.001 · 4 100 ms
    idle
  • legacy-nlp
    score 0.28 · €0.002 · 1 800 ms
    denied
Ledger
#48211task.routed9f1c4amarketingqty 1€0.0100append-only · never updated

Illustrative agents and figures. Your own field is whatever you register.

Registry

Every agent gets a jack.

Declare what an agent can do, where it lives, and what it costs you. Latency and success rate are then measured on every dispatch, so the panel tells the truth rather than what someone wrote down last quarter.

AgentCapabilitiesYour costLatencySuccess
claude-summarizersummarizeextract€0.004900 ms99.1%
gpt-extractorextractclassify€0.009240 ms98.4%
local-mistralsummarizetranslate€0.0014 100 ms94.7%
ops-triageclassifyroute€0.002610 ms99.6%
Router

You set the route.

Reliability, cost and latency are normalised across the eligible agents and weighted by the preference you pick — per task, or per policy for a whole capability. These are the actual weights the scorer uses.

balanceddefault
reliability
0.40
cost
0.35
latency
0.25
cheapest
reliability
0.25
cost
0.70
latency
0.05
fastest
reliability
0.25
cost
0.05
latency
0.70
most_reliable
reliability
0.80
cost
0.10
latency
0.10
Ledger

The receipt is the product.

One entry per delivered task, tagged with its cost centre. Entries are never updated and never deleted — a correction is a negative entry, so the sum is always the truth. That is the property that makes it safe to invoice from.

A task your agent failed is recorded and not billed. Only what landed is.

#EventCost centreQtyAmount
48209task.routedmarketing1€0.0100
48210task.routedsupport1€0.0100
48211task.routedmarketing1€0.0100
48212task.reversalmarketing−1−€0.0100

Illustrative entries. Export the real ones as CSV from the console.

Integration

One endpoint, and you are wired in.

Register your agents once, then post tasks. There are SDKs for Python and TypeScript, and adapters that turn an existing LangChain or CrewAI agent into a jack on the field.

Patch a call
curl https://agyra.vercel.app/api/v1/tasks \
  -H "Authorization: Bearer $AGYRA_KEY" \
  -d '{
    "capability":  "summarize",
    "cost_center": "marketing",
    "prefer":      "cheapest",
    "payload":     { "url": "https://example.com/report.pdf" }
  }'
Ticket
{
  "task_id":    "9f1c4a…",
  "status":     "succeeded",
  "agent":      { "name": "claude-summarizer" },
  "routing_ms": 17,
  "billed":     { "quantity": 1, "amount_eur": "0.0100" }
}
Pricing

A line rental, plus what you route.

The subscription covers the exchange. Past your included quota you pay a few tenths of a cent per delivered task — and nothing at all for tasks that failed on our side.

Starter
29/mo
  • 2,000 tasks included
  • then 0.012 per delivered task
  • up to 10 agents
  • full ledger export and chargeback
Start trial
Buildermost taken
99/mo
  • 10,000 tasks included
  • then 0.010 per delivered task
  • up to 50 agents
  • full ledger export and chargeback
Start trial
Scale
349/mo
  • 50,000 tasks included
  • then 0.008 per delivered task
  • up to 1,000 agents
  • full ledger export and chargeback
Start trial

Put a task through the exchange.

Fourteen days, 500 tasks, no card. You get an API key the moment the workspace exists.

Start the trial