OpenInterpretability
  • Ekbasis
  • Guide
  • Pricing
  • Research
  • Lab
  • Tools
  • Notes
  • Registry
  • Manifesto
Sign inGet API keyAPI key
OpenInterpretability

Ekbasis API: know what an action will do before your agent runs it. From an independent lab for AI-agent safety. Open weights, Apache-2.0.

Get your API key
Ekbasis API
  • Console · sign in
  • Getting started
  • Pricing
  • For agents (agents.md)
  • Results and models
  • Client and CLI
  • Cookbook
The lab
  • About the lab
  • Open tools
  • Observatory
  • ProbeBench
  • InterpScore
  • Academy
Research
  • Manifesto
  • Roadmap
  • Papers & posts
  • Docs
Community
  • GitHub org
  • Notebooks
  • SDK (openinterp)
  • HuggingFace
© 2026 OpenInterpretability — Apache-2.0 for code, CC-BY 4.0 for docs.Built in public.
Back to EkbasisHOSTED API · LIMITS · DATA · VERSIONS

API policies

What the hosted API at openinterp.org/api/v1 limits, keeps and changes. Self-hosting the open weights involves none of this: nothing leaves your machine. Uptime, latency and daily quality, measured from outside: API status.

Rate limits

The hosted API runs on shared GPUs, so there is one limit for all callers together, not a quota per key.

LimitValue
Requests in flight at once10
Input tokens per second, on average≈ 9,000
Burst150,000 input tokens
Wait for a free slot before answering 429up to 20 s
POST /feedback200 calls per account per UTC day

When it is still full after the wait, the API answers 429 with Retry-After: 5. Retry after that many seconds. The Python and TypeScript clients raise CannotJudge on a 429: treat the action as unchecked, never as safe. There are no X-RateLimit-* headers, because a shared limit has no per-key remainder to report.

What we keep, and for how long

Not kept: what you send and what the model answers.The state, the questions, images and the answers are never written anywhere, by the gateway or by the model server: no database, no log file. The model server keeps a prompt's cache in GPU memory only while it serves. For the same reason the hosted API has no sessions.

Kept: metadata about each request, used for billing, the console's usage view and running the service.

WhatKept
Account id and email, last 4 characters of the key, path, HTTP status, input tokens, time in ms, IP address and country, error message, request id, the first 80 characters of the User-Agent, X-Ekbasis-SurfaceA rolling stream of about the last 20,000 requests, all accounts together
The same without IP address or email, in your account’s call list (the console’s usage view)Your last 10,000 calls, up to 400 days
Hourly totals8 days
Daily totals400 days
Which account owns each request id (so only you can send feedback on it)30 days
Feedback you send (POST /feedback): the verdict, your note of up to 500 characters, the call’s metadataThe last 100,000 entries; daily counters 400 days

Questions, or a request to delete your account's records: [email protected].

Models and versions

  • The hosted API serves one model at a time, named by GET /api/v1/v1/models and by the X-Ekbasis-Model header on every answer. Today: ekbasis-27b-int4-v1.3 (weights caiovicentino1/Ekbasis-27B-INT4). Record the header next to the decisions you log if you need to know later which model made them.
  • Before the served model changes, the new version and its scorecard (accuracy, confident errors, calibration, latency and cost on fixed suites) are posted at least 14 days ahead. The change itself goes in the changelog with its date. A security fix may ship at once and is listed the same day.
  • The request and response formats of /v1/systemonefollow the client's semantic versioning: a field is never removed or redefined without a new major version. New fields may appear at any time, so ignore fields you do not know.