On August 4, OpenRouter published a blog post announcing Ori Harness. The post is explicit: Ori is not a new agent. It is a launcher/wrapper layer — ori runs an agent CLI already installed on your PATH, injecting OpenRouter credentials, model settings, and environment variables on your behalf. The official tagline: *"Your workflow does not change."* If the target agent isn't installed, Ori prints the install command.
What it does
- Supported harnesses (4): Claude Code (
ori claude), Codex (ori codex), OpenCode (ori opencode), Hermes (ori hermes). No Cursor, no IDE/GUI tools. - Install:
curl | bash. - Auth:
ori loginuses OAuth PKCE tied to your existing OpenRouter login — no creating or pasting API keys. - Model switching: pass any OpenRouter model ID via
--model; everything after Ori's own flags is passed through verbatim to the agent. - No official open-source repo. The
piratheon/orifound in searches is an unrelated third-party C++ TUI — a name collision; don't cite it. - Ori lives under
/labs/; OpenRouter states it may change or be removed at any time. - Install is
curl | bashonly; no stated native Windows support. - The "nearly half the tokens" claim is OpenRouter's own, with no benchmark or methodology, and applies only to Anthropic models.
- Ori is not a cost-saving smart router: it does no per-request intelligent routing (unlike claude-code-router's task-based splitting). Savings come from manually choosing cheaper models plus org budget caps.
- As of research time: zero independent hands-on reviews.
- Official announcement: https://openrouter.ai/blog/announcements/ori-harness
- Official docs: https://openrouter.ai/docs/guides/ori/harness
- Product page/FAQ: https://openrouter.ai/ori/harness
- Guardrails: https://openrouter.ai/docs/guides/features/guardrails
Why it matters: gateway degradation, made explicit
The most substantive information is that Ori surfaces the cost of manual gateway wiring. OpenRouter's own list of 13 environment variables for hooking Claude Code to the gateway includes the critical one: ENABLE_TOOL_SEARCH. Enabling it for Anthropic models reportedly saves nearly half of system prompt tokens.
Modern Claude models only enable tool search by default when they detect a first-party harness — this is the root cause of performance loss when routing through third-party gateways: with it off, Claude burns more tokens, takes more turns, and calls the Todo tool more often. ori claude also detects --model and switches to that model's optimal configuration.
Cost control
Ori inherits OpenRouter org Guardrails: USD budget caps (daily/weekly/monthly reset), model allowlists, provider allowlists, ZDR, and per-workspace permissions — enforced server-side on every request, returning 403 on violation. Billing is folded into Usage Accounting; there is no separate Ori bill.
Strategic read
Reproducible multi-model configuration has landed: one command + OAuth replaces config scattered across shell profiles and settings.json, making team setups copyable and auditable. The product-line signal is clear — Ori Eval (model selection benchmarking) shipped August 3, Ori Harness on August 4, both sharing one install.sh. OpenRouter is moving up from "routing API" toward being the entry point for AI coding workflows (credentials, model selection, governance, billing in one place), raising switching costs.