[Question] Support for CLI-based executors (claude-cli) vs. Direct API Integration

[qitma]

I am exploring the integration of Evonic with claude-cli to execute headless prompts. Specifically, I am curious about the intended flow for Super Agents:

Is it possible to have Evonic trigger a local CLI command (e.g., claude-cli -p "command") and capture the output directly as an agentic action?
Or, does Evonic require a direct LLM API key integration by design, where the CLI would instead be defined as a "Skill" or tool for the agent to call?
I am trying to determine if I can bypass direct API costs by leveraging a locally authenticated CLI session.

Evonic --> super agent --> trigger claude-cli -p "some comand" --> return output

Thanks mas robin.

[binsarjr]

Hi @qitma 👋

In my view Evonic isn't really ready to take on first-class integration with CLI-based AI tools (claude-cli and friends) yet. The main reason is maintenance cost — these CLIs change fast and frequently, and wiring one in as an executor means committing to chase every breaking change in their flags, auth flow, and output format. That's a lot of surface area to keep stable over time.

For your specific goal — avoiding direct API costs by reusing an authenticated session — there are a couple of practical paths today:

1. OpenAI-compatible proxy — expose your Claude subscription behind an OpenAI-compatible endpoint using a router/proxy like 9router, then point Evonic at that endpoint as a regular provider. That's the setup I've been running myself.
2. Skill + dedicated agent — define claude -p as a Skill/tool and wire it to a dedicated agent whose whole purpose is to always invoke that claude -p call. So instead of Evonic owning the CLI integration, the agent is instructed to route everything through claude -p.

Just go in with eyes open on either path: they carry their own trade-offs (reliability, and it sits outside the officially supported usage), so weigh the risks before relying on them.

More broadly, I think Evonic would benefit from solidifying its core fundamentals first before expanding into a lot of external tool integrations. Otherwise the integration surface tends to grow faster than it can realistically be maintained. Honestly this is partly just a personal concern of mine — a lot of the other agentic swarms out there feel way too bloated, and I'd hate to see Evonic head down that same path.

That said, this is just my take — curious how others see it. Do you think CLI executors are worth the maintenance overhead, or would an API/proxy approach cover most use cases? Happy to discuss further.

[anvie]

Yes, as @binsarjr said, we're currently focusing on the fundamentals first. We recommend using 9router for your case, but be cautious of Anthropic's restrictions