Evaluate
Evaluate Saylek without an account.
This page exists because a reviewer told us the site explained the idea well and gave them no way to establish that the thing works. So: what the system is, where the trust boundaries fall, what it runs on, what the API looks like, what provenance you can check, and what is still missing. Nothing below needs a session, and the last section is the list of things you cannot verify yet.
1. What Saylek is, mechanically
You install a daemon on your own machine. It serves AI models locally, and exposes an OpenAI-compatible API on loopback for whatever you point at it.
If you join a Circle, a group of members who share machines with each other, two further things become possible. When your own machine cannot serve a request, another member’s machine can serve it for you. And if you choose to arm it, your machine can serve theirs while your GPU is idle.
There is no billing anywhere in it. Contributing capacity earns nothing and drawing on the pool costs nothing; what circulates is a record of who served what, not a balance anyone can cash.
2. The request path, and where the boundaries are
This is the part worth reading closely, because the honest version is less comfortable than the pitch. A request that your own machine cannot serve crosses three boundaries:
Your machine
The daemon, your keys, your models and your receipts, under ~/.saylek. The local API binds loopback and is not authenticated: it reads an Authorization header and does not check it, so anything that can reach the port can use it. Widening the bind is possible and is documented as the hazard it is.
Saylek's registry
Your prompt leaves your device as text, not as a hash or a summary, and reaches the serving machine by way of our registry, which routes it. So we handle it too. The outbound queue is trimmed by count, around 10,000 entries, and never by age, which means a quiet queue can hold a prompt indefinitely. The completion coming back is cleared once delivered, with a ten-minute backstop.
The other member's machine
Their machine reads your prompt in the clear, because that is how it computes an answer. There is no end-to-end encryption and no enclave, and we are not going to imply otherwise. What they do not get is your identity: your machine presents a key fingerprint. It is stable, so a Host can tell that many requests came from the same member without knowing which member.
The control that stops this is not joining a Circle, or setting local-only mode, which fails a request rather than sending it. Neither overrides a proxy upstream you configured yourself. The egress page is the full version, and terms section 3 is the same disclosure in the agreement itself.
The internal architecture notes and the STRIDE decomposition stay in-repo. Everything on this page is already published across the terms, the privacy notice and the docs; what is new is having it in one place.
3. What it runs on
Operating systems
Released binaries target Linux x86_64 and macOS arm64 (Apple Silicon). The Linux build is cross-compiled to a glibc 2.31 floor so it runs on older distributions. There is no Windows build.
Acceleration
cuda for NVIDIA, metal for Apple Silicon, and a portable cpu backend. If you build from source with no backend selected it compiles and then answers 503 on the first call, which surprises people, so it is worth knowing before you file it as a bug.
Models and runtimes
The daemon serves models discovered on the machine and reuses what you already have rather than insisting on its own store. It can also front proxy upstreams you register yourself. saylek models lists what a machine can serve now; saylek models --circle lists what the Circle can reach, which is a different question.
4. The API surface
Three base URLs. The local daemon at http://127.0.0.1:8443/v1 with no auth, a hosted OpenAI-dialect surface at https://api.saylek.com/v1 with a bearer key, and a hosted Anthropic-dialect surface at https://api.saylek.com which takes no /v1 suffix.
Implemented: GET /v1/models, POST /v1/chat/completions, POST /v1/embeddings, POST /v1/audio/transcriptions, and POST /v1/messages on the hosted Anthropic surface. Two carry caveats worth having before you test: /v1/responses works only against a proxy upstream you configured, and /v1/completions is not implemented and answers 501 rather than 404.
Bodies are capped at 32 MiB (413 over that). Rate limiting is on by default at 100 requests per second, burst 200, returning 429 with a Retry-After. The API reference carries the error table.
5. Provenance: what you can check, and what you cannot yet
Every served request produces a signed receipt naming the machine that served it. Both sides keep a copy, and the verification is offline: the point of the design is that you do not have to ask us whether a receipt is genuine.
The verifier is a separate program from the daemon, deliberately. saylek-verify is built from the receipt core alone and a test in the repository fails the build if it ever links daemon code, so what checks a receipt cannot also be what produced it. It ships signed for both released targets. If you have the daemon installed, saylek receipt show <id> --verify checks one receipt.
Now the part a reviewer actually needs, and the answer is no. There is currently no way for you to do this without an account. The verifier is published behind the same token-gated path as the daemon binaries, so the download link returns 404 to an anonymous request, and we have not published a sample receipt for you to run it against. The public release manifest tells you the current version and nothing else.
That means the strongest claim on this site is, for you today, still asserted rather than demonstrated. We would rather write that sentence than let the architecture stand in for the evidence. An anonymously downloadable verifier plus an anonymised real receipt is a tracked gap, not an oversight.
6. Security, privacy, and what is kept
Three answers most evaluators want before they read anything else.
Retention
Nothing on a machine is deleted automatically. Receipts and logs on your machine and on a Host’s persist until someone runs saylek wipe. There is no TTL and no cleanup job. Deleting an account is a 90-day soft delete after which it stops being recoverable; the record is not purged or anonymised today.
What we do not do
We do not train on prompts, do not sell information about members, and do not read prompts for analytics or moderation. Nothing inspects or filters request content anywhere in the system, which is a disclosure rather than a reassurance: the Acceptable Use Policy explains what that means for both sides.
Outside parties
Cloudflare (site, account records, bot check) and Resend (email delivery). Everything else runs on Saylek’s own servers or on a member’s machine.
Full versions: Privacy, Security, Terms, Acceptable Use. Vulnerability reports go to the disclosure policy, which states scope and safe harbour, and /.well-known/security.txt is the machine-readable pointer to it.
7. Availability and limitations
Saylek is in public beta. Anyone can create an account: registration takes an email address and a bot check, with no invitation and no allowlist in front of it. What is still gated is the binaries, by the download token section 5 describes, and the site is excluded from search indexing while we harden it. Stated plainly, so you can calibrate:
No availability guarantee
The machine that would have served you belongs to someone who can switch it off mid-request and owes you nothing. That is the model working, not the model failing.
Inference only
Training, fine-tuning, running containers, and logging into someone’s box are not what this opens up.
Priority is bounded, not instant
A Host’s own work takes priority on their machine and an in-flight guest request is dropped when owner work arrives, but the interruption lands at the next step of the work rather than immediately.
No acceptance step yet
Terms are shown and the version in force is recorded at sign-in, rather than captured by a separate consent step. The terms say so themselves, under section 11.
8. What is missing from this page
Two things a reviewer asked for that are not here, named rather than omitted:
A recording of the product working
There is no two-minute capture of a real session. Until there is, this page is documentation rather than demonstration, and you should read it that way.
A receipt you can verify yourself
Covered in section 5. The verifier exists and is not anonymously downloadable, and no sample receipt is published.
Both are tracked. If you are evaluating Saylek and either one is what would settle it for you, say so at Contact and it moves up the list.
Written for a technical evaluator with no account and no reason to take our word for anything. Every factual claim here is grounded in the code or configuration that produces it, and where we could not ground something it is on the missing list rather than on the page.