Documentation

Get started with Sageros

Sageros is a drop-in gateway. You keep your existing agent code and simply route model calls through Sageros instead of calling providers directly.

Early access - what works today. Live now: the one-line cloud gateway and inline PII redaction. In development and documented here as the intended design: session supervision, semantic caching, human-in-the-loop approvals, and policies. Sections below marked in development describe where we are headed, not shipped behavior.

Quickstart

Change one line: point your model client's base URL at your Sageros gateway and pass your session id. That's it - supervision, caching, and policy enforcement turn on automatically.

# point your OpenAI-compatible client at Sageros
base_url = "https://gateway.sageros.com/v1"
headers = { "X-Sageros-Session": "sess_a1f3c9" }

# your agent code stays exactly the same
resp = client.chat.completions.create(model="gpt-4o", messages=msgs)

Open the dashboard and you will see the session appear in real time.

Core concepts

  • Session: one agent task, grouped under a single sageros_session_id, containing many calls.
  • Anti-loop breaker: automatically halts an agent that repeats the same call 5+ times, preventing runaway spend.
  • JIT approval: pauses a destructive tool call and asks a human to Approve or Deny before it runs.
  • Semantic cache: serves a stored answer when a new prompt is close enough to a previous one.

Supervising sessions

Every session shows a live timeline of its calls - LLM calls, tool calls, cache hits, and blocked actions. From the dashboard you can watch a session, arm the kill-switch, or step in on an approval. Nothing destructive happens without passing your rules first.

Semantic caching

Set a similarity threshold (for example 0.92). When a new prompt scores above it against a cached prompt, Sageros returns the cached answer and skips the model entirely - cutting both cost and latency. You can tune the threshold and cache lifetime per environment.

Policies

Partly live Inline PII redaction runs today. The broader policy model - global switches to block EU data egress for GDPR or require approval on destructive operations across every agent - is in development, not yet enforced.

Ready to see it live?
Open the dashboard and fire a simulation.
Launch Dashboard →