v0.1 · research prototype · Apache 2.0
The conversation
is the computation.
Colloq starts with one typed conversation instead of a client program, a server program and a protocol file that drift apart. The compiler projects that conversation into one checked endpoint per role.
$ colloq check examples/route.colloqconv.json
valid conversation example.route (2 roles, 11 states)
$ colloq compile examples/route.colloqconv.json
wrote build/route.colloqplan.json
conversation sha256:92c198d0…
plan sha256:016d5975…
$ colloq run-plan build/route.colloqplan.json --transport quic --wire compact
✓ client and server report the same semantic traceOne source of truth
Write the agreement once.
Every server gets its half.
A conversation names the roles, the messages between them, the choices, the loops, the cancellation and the failures. The compiler derives a local state machine for each role, so two servers cannot disagree about the protocol.
conversation Route(request: Request) {
roles gateway, router
gateway -> router: request within 20ms
choice router {
cached { router -> gateway: CachedAnswer; end }
infer { router -> gateway: ModelAnswer; end }
review { router -> gateway: ReviewTicket; end }
}
}-
Typed choices, not guesses
A branch label is part of the type. A role can only send a transition the plan declares, and the receiver checks it before it advances.
-
Failure is in the language
A timeout or a reset is a declared transition to a declared state, not a generic exception that each service handles differently.
-
One identity across the cluster
The conversation and the compiled plan are content addressed. Both peers exchange those identities before frame zero, and a mismatch stops there.
Transports
The meaning survives the wire.
The same plan runs over shared memory, TCP, authenticated QUIC and Iroh. Every transport reconstructs the same semantic trace, so a local run and a two-host run mean the same thing.
Jev
A model may pick the branch.
It may not break the protocol.
Bind a choice state to a TypeSafe Jev question and the model selects a declared label with a calibrated confidence. Below your threshold, Colloq takes the escalation branch you declared. A service error is a typed failure, never a silent guess.
| Request | Jev picks | Confidence | Colloq sends |
|---|---|---|---|
| “What are your opening hours?” | cached | 0.99 | cached |
| “How does endpoint projection work?” | infer | 0.94 | review escalated |
| “Charged twice, I want a refund.” | review | 1.00 | review |
Live run, threshold 0.95. Three requests is a smoke test, not an accuracy measurement.
Measured, not claimed
Here is where Colloq loses.
A research language that only publishes its wins is marketing. These numbers come from the checked-in benchmark on an Apple M4, 500 iterations, three tokens.
The compact wire is still JSON. The target of 1.25× is not met. There is no hosted product, no scheduler and no SLA. What exists is a checker, a compiler, a reference runtime over four transports, an Automerge draft gate and 76 tests.
Read the benchmarkStart with the two-node run.
Clone it, compile one conversation, then run the same plan across two processes with mutual authentication.
$ git clone https://github.com/Vantar-AI/colloq
$ cargo run -- check examples/generate.colloqconv.json
$ cargo run -- bootstrap-two-node --out build/two-node