Book Nok for four tomorrow at 7pm
Try a request
Pick something a user might say. Bluwing returns the tool, the arguments, and how long it took.
Thetool-callingengineforAIagents.Therighttool,realarguments,inmilliseconds.
The tool-calling engine for AI agents
Bluwing decides which of your tools to call and fills every argument from what the user actually said. It speaks OpenAI’s format, answers in about 60 ms, and costs a fraction of a cent a turn.
200 free credits. No card needed.
Example 1 of 4: “Book Nok for four people tomorrow at 7pm”. Bluwing calls book_table with restaurant Nok, party size 4 and tomorrow at 19:00, each taken from the sentence.
Slowed down so you can follow it. The clock shows the engine’s target time.
Why it's cheaper
An LLM rereads your whole tool list on every turn, and you pay for each of those tokens. Bluwing charges one credit per turn, however many tools you have.
Times are engineering targets on 8 vCPUs with 2,000 tools, not independent benchmarks. Bluwing chooses tools and fills arguments; your app or LLM still writes the reply.
Features
Bluwing does one job: decide which tool to call and with what. These tiles are it doing that job. Try the big one.
Book Nok for four tomorrow at 7pm
Pick something a user might say. Bluwing returns the tool, the arguments, and how long it took.
About 60 ms a turn, against one to three seconds for an LLM.
Book Nok for four people tomorrow at 7pm
Values come from what the user said or facts you pass in. Nothing is made up.
Book Nok tomorrow at 7pm
Sure. For how many people?
needs_input40 msFour
Booked Nok for 4, tomorrow at 19:00.
book_table44 msA missing detail comes back by name, so your app can ask.
“What’s the capital of Peru?”
tool_calls: []7 ms
Small talk and questions get no tool call, so nothing fires by accident.
2,000tools in your toolset
A tree narrows 2,000 tools to a dozen before it chooses, so more tools don't mean worse picks.
from openai import OpenAIclient = OpenAI()client = OpenAI(base_url="https://api.bluwing.si/v1")res = client.chat.completions.create( model="gpt-4o", model="brick:support-desk@live", messages=messages, tools=tools,)
Keep your OpenAI client. Change the base URL and the model name.
Test a new version on staging, then move live to it. Rolling back is one call.
How it works
Upload the tool definitions you already have. From then on, every request is a normal tool-calling request.
{ "name": "support-desk", "tools": [{ "type": "function", "function": { "name": "gmail_search", "description": "Search Gmail by text, sender and date.", "parameters": { "type": "object", "properties": { "query": { "type": "string" }, "sender": { "type": "string" } }, "required": ["query"] } } }]}
Your request is checked, 2,000 tools are narrowed to the one that fits, and each argument is filled from the conversation. No LLM is called on the way.
Promises
An agent that calls tools has to be predictable. Bluwing is built around nine rules, and each one is checked by its own tests.
Rule 01 of 9
Every value comes from the request, the chat, your facts or an earlier tool result.
What it leaves to you: writing the reply to the user, and running the tool itself. Running tools on our servers is planned for v1.0.
Pricing
Every plan runs the full engine. Move the ruler to your volume: each plan shows what that month would cost, and the cheapest one lights up.
Calls a month
Cheapest: Plan A at $0.00 a month. An LLM would cost about $0.00.
You keep $0.00 a month against LLM tool calling.
Try the full engine, then pay as you go.
$0.00/ mo
All top-ups, after 200 free creditsFor side projects and prototypes.
$0.00/ mo
0% of the 500M included creditsFor startups with agents in production.
$0.00/ mo
0% of the 1B included creditsFor platforms running fleets of agents.
$0.00/ mo
0% of the 2.2B included credits| Limit | Free | Plan A | Plan S | Plan S+ |
|---|---|---|---|---|
| Monthly price | $0 | $4.50 | $9.00 | $19.00 |
| Tokens included | 5M at signup | 500M / mo | 1B / mo | 2.2B / mo |
| Cost per 1M tokens | $0.015 / 1M | $0.015 / 1M | $0.015 / 1M | $0.015 / 1M |
| Top-up per 100M tokens | $1.50 | $1.50 | $1.50 | $1.50 |
| Execution speed | Standard | Accelerated | Ultra-fast | Real-time |
| Priority queue | Standard queue | Fast-path lane | High priority lane | Dedicated zero-wait lane |
| Rate limit | 10 RPS | 50 RPS | 200 RPS | 1000 RPS |
| Requests in flight | 4 concurrent | 16 concurrent | 64 concurrent | 256 concurrent |
| Active toolsets | Unlimited | Unlimited | Unlimited | Unlimited |
| Edits and versions | Instant live rollouts | Instant live rollouts | Instant live rollouts | Instant live rollouts |
Prices in USD. Dodo Payments handles checkout and collects sales tax and VAT. Estimates leave out toolset builds and round top-ups up to the next 100 credits. The LLM figure uses the assumptions shown above.
Start on the free credits. Pay only when the volume is real.