Status
Service status
Antecedint is early. This page says what is deployed and what its limits are, and it is maintained by hand rather than driven by a monitor.
What is deployed
MCP server
mcp.antecedint.comSingle instance behind a load balancer in us-east-1. Best effort, no availability commitment.
Website
antecedint.comStatic site on a CDN. Not affected by API deployments.
What to expect
One instance, one region
The API and its inference service run on a single machine in us-east-1. A failure there is an outage, not a failover, and a deployment is a short interruption to new requests. We publish this rather than imply redundancy that is not there.
Deployments drain, they do not cut
A restart stops accepting new requests, then waits up to 210 seconds for work already in progress to finish. A tool call running when a deploy starts completes normally. A call that starts during the window gets a 503, and retrying it works.
Reranking is slower on this hardware
The reranker runs on CPU, not GPU. Prior-art searches with rerank enabled take tens of seconds, and the document length the model reads is capped to keep them inside the tool timeout. Searches without reranking are unaffected.
No latency percentiles yet
We have not published p50 or p99 numbers for the tool calls. The per-tool timeouts on the documentation page are ceilings, not expectations.
No SLA at launch
There is no uptime guarantee and no support-response commitment on the public endpoint today. If you need either, they are something to write into an agreement rather than something to read off this page. The architecture is stateless, so adding a second instance is a configuration change; when that happens, and when there is monitoring worth publishing, this page will change with it.
Reporting a problem
Email accelerate@antecedint.com with the time, the tool you called, and the HTTP status or JSON-RPC error code you got back. If you have the key identifier (the ak_live_ prefix and the eight characters after it, not the whole key), include it: we can find your requests in the logs from that alone.