There are two ways to put TELOS in front of your model API. They run the same TELOS pipeline, the same state accumulation, and the same cache_control injection — the only differences are the process boundary, error handling, and streaming.

Path B · HTTP reverse proxy

Recommended. Zero code changes — set ANTHROPIC_BASE_URL to the local gateway. This is what telos init configures.

Path A · SDK transport

In-process. Swap the LLM client for the TELOS transport; the duck interface is identical.

​
How they compare

Path B · ProxyPath A · SDK transport
IntegrationANTHROPIC_BASE_URL env var, zero code changeswap the client object
Process modelout-of-process (separate gateway)in-process
StreamingSSE-aware reverse proxydirect SDK streaming
Session statekeyed by session_id in an LRU registryone BridgeSessionState per transport instance
Best forany harness that respects ANTHROPIC_BASE_URL / OPENAI_BASE_URLapps you control the source of
Failure modenon-strict by default: degrades to passthroughraises in-process
Both paths share the same pure function — proxy/pipeline.py:process_anthropic_request(raw, ...) splits out parse → bridge → emit, eliminating any wire drift between the two.

​
A request, end to end (Path B, mode=both)

​
Pick one

  • You run a CLI agent (Claude Code, Codex, OpenClaw, Hermes). Use Path B — run telos init and you’re done. See Harness integration.
  • You’re building your own agent in Python. Either works; Path A keeps everything in-process. See SDK transport.