That said, while I'm hardly a fan of MCP (judge for yourself by reviewing my previous comments on the matter), at least its security model was standardised around OAuth, which in my opinion is a good thing, albeit with a few small issues.
I personally prefer CLIs, but their security is in fact worse. A lot worse! Sure, we can now store API keys in a vault, but it's not like you can rotate or expire them easily. Plus, the security model around APIs is based on path-based rules, which aren't very effective given that most services use REST-style APIs. This is even worse for GraphQL, JSON-RPC, and similar protocols.
It is backwards. I bet we will move from CLIs to something else in about 3-6 months.
As long as the fake keys are known, they can be mapped directly to the real key with the endpoint in OneCLI to exfiltrate the data and you don't need to leak any keys anyway.
The correct solution is that there should be no sort of keys in the VM / Container in the first place.
> It is backwards. I bet we will move from CLIs to something else in about 3-6 months.
The hype around CLIs is just as unfounded as was MCPs and made no-sense just like OpenClaw did. Other than hosting providers almost no-one is making money from OpenClaw and from its use-cases; which is just wasting tokens.
We'll move on to the next shiny vibe-coded thing because someone else on X said so.
MCP has plenty of problems, but standardising on OAuth was one of the better calls. Expiry, scopes, rotation, delegated access, all much better than the usual CLI pattern of long-lived API keys. The CLI story there is still pretty rough.
And once the policy model is host/path matching, GraphQL and JSON-RPC become awkward immediately unless the proxy starts understanding payload semantics.
My bet would be OpenAPI specs. The model will think its calling a cli but we intercept the tool call and proxy it with the oauth credentials.
There are some implementations already out there in open web ui and bionic gpt.
1. Full secret-memory isolation whereby an agent with root privileges can't exfilrate. Let's assume my agent is prompt injected to write a full-permissions script to spin up OneCli, modify the docker container, log all of the requests w/ secrets to a file outside the container, exfiltrate.
2. An intent layer on top of agents that models "you have access to my gmail (authN) but you can only act on emails where you are a participant". This would be more similar to universal RBAC between agent ↔ mcp etc.
I've been building on [2] for a while now using signed tokens expressing intent.
On (1), the agent runs in its own container where OneCLI doesn't exist. It can't spin up OneCLI or access its process because it's completely isolated from it. The agent only ever sees placeholder tokens, the real secrets live in a separate container it has no way to reach.
On (2), we actually address this with OneCLI Rules, deterministic constraints enforced at the proxy level before a request ever hits the API. So the agent doesn't need to "behave", it just can't do what the rules don't allow. Would love to hear more about your signed tokens approach.
I still wouldn't give to any claw access to my mail accounts, but it is a step in the good direction.
I love how NanoClaw is aggregating the effort of making personal assistants more secure.
Good job!
I also do not understand the uses right now. Are we just grasping at usefulness ?
Also, attaching an LLM with my raindrop.io and Todoist credentials to cron is fun. Haven't got the kinks worked out, yet, but it's pretty incredible how much data-shifting I can do now. Saved me a lot of clicks.
I had it make a daily aviation weather brief for a private airpark. It uses METAR, outdoor IP cameras I have including one that looks at a windsock and another that looks at the runway surface, and a local weatherstation. It sends me a text message with all of that information aggregated into "It's going to be really windy this afternoon, visibility is high, but there is ice on the runway surface", that sort of thing.
The thing is, all I had to do is point it to a few endpoints and it wrote the entire script and set up a cron for me. I just gave it a few paragraphs of instructions and it wrote, then deployed the rest.
The other day, there was a post here about a new TTS model. I wanted to try it out, so I gave my claw the github URL, and it pulled everything down and had it running without any effort on my part. Then it sent me a few audio messages on discord to try.
When I'm away from home, I can text it to say "what's going on at home" and it will turn on the lights around the house, grab a frame from each camera turn lights back off, and give me a quick report. I didn't have to do any work other than tell it I wanted that skill.
I also have a group chat with some friends on signal that's hilarious. It roasts us, gives us reminders, lets us know about books we might be interested in, that sort of thing. It's really fun.
I've set it up with it's own mailbox, and a git token to make PRs etc. So far I've set up a few automations (check thing X and message me if Y) but the combination of enough "intelligence" to be able to triage/suggest solutions, messaging via a standard messaging app, and a sandboxed environment for code execution, all packaged as "this is your helpful assistant" is fun.
In theory it's nothing I couldn't do with Claude Code + some integrations, but having all of that out of the box + setting the expectation that this is legitimately a helpful assistant you can message and email with any mad request you have does shift the way I see it.
Though yes, the more fun you have with it, the more of a security threat the whole thing becomes, and it's slightly terrifying. I briefly considered giving it view only access to my emails and decided that was just too high risk. But treat it as a vaguely clueless but not incompetent intern and works?
1. monitoring anything online and giving me a summary when something changes
2. contribute or make edits to any online forum (where TOS allows it)
3. Giving it access to any API / cli gives you a natural language interface to that service.
4. Memory / notes retrieval. It can search through its discussion / thought history and answer detailed questions about what happened in the past.
5. Any standard GPT cases but it has a much more specific memory of who you are and what you might be interested in.
6. If you ever want to add capabilities you just tell it to add a new skill to do xyz and it just does it.
NanoClaw is pretty interesting to me. It's a really small piece of software and a different category of software than we've seen before, one that you are expected to customize by editing it yourself (via Claude Code).
For that reason alone I find it interesting to play around with. I don't find it actually useful at the moment.
I find it interesting how European and American (especially on HN) are so luddite about AI and OpenClaw, but Chinese, Indian, and Israeli developers both domestically and in the diaspora have been adopting fairly rapidly.
And then people wonder why the center of gravity for a number of engineering subfields is slowly shifting.
Nice idea, but it will not work. Agents are so resourceful and determined, they will find that weird call which can delete all emails with one request (/delete?filter=*)