A multi-agent system delivered as one model.
Behind the single model string semar, a coordinated pool of models answers as one voice. Point an existing OpenAI or Anthropic SDK at one endpoint, change nothing else, and keep your code.
invitation-only · both wire protocols · one model string
Two admissions before anything else. Answers typically land in 15–90 seconds; Dalang is a fixed-quality answer engine, not a sub-second chat model. And pricing is published at general availability, not before. Nothing on this page is a price.
Integration is a base_url change.
Dalang speaks both the OpenAI and the Anthropic wire protocols natively: request and response shapes, streaming, and tool calling included.
OpenAI SDKs and clients
Anthropic SDKs and clients
Tool calling passes through on both surfaces. Your tools go in as content; the answer comes back as tool_calls or tool_use blocks on the matching surface. That is what lets agentic clients like Claude Code point at Dalang unchanged.
The model
One model string. No tier to pick.
semar is named for the wayang’s wise guardian, the humble figure who quietly carries the whole performance.
One string, priced as one
There is no model zoo and no tier menu. Easy requests stay fast and cheap; hard ones earn more coordination automatically. Adding agents to the pool does not multiply your bill: billing counts the tokens you can see, at one blended rate.
Opaque on purpose
Dalang will not tell you which models answered or how they were coordinated. The routing is the product. What you get instead of a peek behind the screen is the measurement: an eval harness built before the gateway itself.
Measured, not imported.
The category habit is to quote other labs’ published benchmark numbers next to your own. Dalang’s harness refuses imported numbers.
- Baselines re-run
- Every baseline model is re-run inside the same harness as the policy that challenges it. Same tasks, same sandboxes, same cost meter. No number is copied from a press release.
- A control for honesty
- A scaffold control (one model plus a single verifier pass, no orchestration) separates genuine coordination gains from simply spending more compute.
- The harness came first
- The eval harness was built before the gateway, on purpose: a coordination claim with no measurement underneath it is marketing.
- Review trail
- Phases shipped through external Codex review; the serving core carries ~466 passing unit tests, verified live against its production database.
What is not here: published benchmark figures. The internal coding suite is 8 dev + 9 test seed tasks. It is not SWE-bench Verified and will never be presented as such. When numbers are cleared for publication, they will come with the harness that produced them.
Questions a skeptic would ask
- Why would I accept 15–90 seconds?
- Because you are not buying latency, you are buying a considered answer. For interactive chat, use a chat model; point Dalang at the questions where quality is worth the wait.
- What happens to my existing code?
- Nothing. If it speaks the OpenAI or Anthropic protocol, it speaks Dalang: change base_url and the key, keep the SDK, keep the tool definitions.
- Can I choose which models are in the pool?
- No. The pool and its coordination are Dalang’s job, and they are deliberately not caller-visible. If a policy change makes answers worse, the harness catches it before you do.
- Why invitation-only?
- Because capacity and support are one person deep, and honest onboarding beats a public signup that underdelivers. Tell us what you are building; invitations go where the fit is real.
Request access
Dalang is invitation-only. Leave an email and you will hear back from the person who built it.
Prefer to write more? Use the contact page