Lavan

Fan-girling over LLM security with @unrealKushagra at yesterday’s AI Tinkerers meetup in Boston

@RhysSullivan x.com/lavanpuri/stat… just asked for you bro

"what would you do if you had a million dollars in anthropic tokens?" "i'll tell you what i'd do man, two fable agents at the same time man" "that's it?" "damn straight, i've always wanted to run two fable agents at the same time"

@AnthropicAI what time tomorrow though? asking for a friend.

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon. We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

im late to the loops discourse but from what i'm seeing it's mostly about creating a loop from your asshole back into your own mouth?

@RhysSullivan But computer use for oauth credentials generation feels like a hack which can break anytime. Most of the flows need human intervention which is a deliberate design decision.

@RhysSullivan - Couple of mac minis added as deployment options in codex and cursor as background agents - All source controlled with a shared-mcp-servers repo - Have global config file per machine that contains the tokens with correct access control.

I've got questions for the people having agents write software for their team: - Where do you deploy it? - How are you giving it access to integrations? - How are you managing access controls? Diving deep on this topic atm, if you're a company struggling with this let me know

Limit your rage prompts folks, it can quit on you

@kushagra_agr Me visiting my own vibe wall for the 75th time today 😎

guyssss, we just crossed 1000 uploaded images on vibewtf.com, so happy!! 🥹

@_nishankj Sounds good, but if there are issues in the intermediate steps in the script running in V8, the model might fall back to the same individual tool call tendencies. Definitely saves tokens for straightforward queries.

@RhysSullivan Doing something very similar with getsesame.dev , the platform specific actions is a very nice use case

The vision I have for Executor is to be able to be able to replace how agents use CLIs and MCPs today with a significantly better primitive to let you call anything. I'm still figuring out the concise version of what Executor is, but for those that want the long explanation: Executor doesn't care what you add. MCP, OpenAPI, GraphQL, Custom sources, it just represents them as a tool name, input schema, output schema under the hood. This lets you then call them however you want. Today that's implemented as a code mode MCP, but it could also be called via an Executor CLI, or dropping in a native client. Also, because of this representation, they can be called via gen UI dashboards your agent writes, they can also be used in reusable workflows. Teams: You're able to set up all of your sources once and now every member of your team has access to them. It also supports per user credentials vs global credentials. Each member of your team shouldn't have to figure out how to set up your agentic environment, you shouldn't have to toggle on / off MCPs while working with your agents. Your agents should be able to access all of your company resources in a way that isn't scary. Sandboxing: Executor executes the tool calls in a JavaScript sandbox, the agent isn't able to access any of the secret values, those are done via proxy calls. Destructive Actions: Executor preserves the semantics of what it imports from i.e GET/DELETE for OpenAPI, destructiveHint for MCP, mutation for GraphQL. This then informs what the agent can auto run / not auto run, letting it work autonomously without removing you from the loop. It's all open source and built on the same SDK that's published to npm. Runs locally, will be self hostable in the cloud, there's so much to build and I'm really excited about it

this is super cool, gonna embed this in my workflow!

@icanvardar You don’t have to anymore! getsesame.dev

DO NOT SHARE YOUR SECRETS WITH AI!

@kushagra_agr Letsss goooo!!

Alright bois, wish me luck!

@CiprianiRanieri getsesame.dev

Builders only. Drop your product. No pitch. Just link

@audiencon Sesame lets you approve exactly what your AI agent can do — which endpoint, which method, what scope, for how long — and revoke that access in one click. getsesame.dev

drop your project i’m boosting builders tonight 👇

@hthieblot Sesame lets you approve exactly what your AI agent can do — which endpoint, which method, what scope, for how long — and revoke that access in one click. getsesame.dev

Describe your product in exactly one sentence. No buzzwords, no fluff, just the core value. If I can’t understand your business in ten seconds, I’m not investing. Hit me & i'll be in your DMs

Try it at getsesame.dev 5 minute install.

Stop pasting API keys into agent prompts or memory files. Your AI agent needs secrets. It doesn't need to see them. Your agent tries to refund a customer. Sesame denies it. Only read access was authorized. Brokered. Audited. Revoked in one click. 70+ providers, bring your own vault.

@kushagra_agr 27 and cracked! Keep building, idea machine!