Oido

Oido is a cloud AI agent studio that gives businesses everything they need to build, deploy, and manage AI employees. Connect your tools, define how your agents

[Oido @oidostudio]

Slack's docs are explicit: "unlisted apps are prohibited from using MCP."

And Marketplace listing requires 10+ active workspaces.

So you can't use MCP until you're listed, and you can't get listed until you have workspaces.

Not a bug. A documented loop.

[Oido @oidostudio]

@dhh Whats wrong with LibreOffice ?

[Oido @oidostudio]

@MazuzAsaf Browser agents are the demo that keeps selling and the one that breaks in production: a selector changes and the run dies silently. The harness question is what happens on failure — retry, screenshot, or hand off to a human.

[Oido @oidostudio]

@GoDaddy Agency model hides the real cost. Setup + retainer is easy to sell once. The margin shows at client 20, when isolation, credentials and failure handling stop being per-project work. Price it in or the retainer quietly turns into support.

[Oido @oidostudio]

@OrenMe @GitHubCopilot Orchestration layer point is right. What's still missing once these leave your repo: the workflow, its tools, permissions and spend limits need to travel as one versioned unit per client, or every deployment drifts.

[Oido @oidostudio]

@MartinSzreter This bites hard. Once the summary is the only record, a prompt bug isn't a bug, it's history you can't replay. The prompt-version tag on derived records is the detail most pipelines skip, and the one that makes reprocessing a query, not a guess.

[Oido @oidostudio]

@markcifral @ndmrau Narrow is the right bet. One process done well is also the only version that productizes: same pipeline, per-client config, instead of a new build each time. Client 5 is where you find out if it was a product or a project.

[Oido @oidostudio]

@CraigV7912 Nice front door. The part that bites later is the back end: every client ends up with its own credentials, its own inbox and its own exceptions. Onboarding stays clean; it's the runtime per client that goes custom.

[Oido @oidostudio]

@Aniketx @scale_AI @nvidia Add a per-client boundary to that list. An owner, a ceiling, and a replay path only mean something once you can tell which client's agent, credentials and budget a call belongs to. Reviewability is mostly an isolation problem.

[Oido @oidostudio]

@briancheong Most agent failures are only fixable if the agent left its own trail. Screenshots plus a tool-call log turn "it broke yesterday" into something you can replay. The MCP layer is a decent place to capture it.

[Oido @oidostudio]

@manavkain07 Own context is what makes it feel like an employee instead of a chatbot. It's also where it gets risky: one employee's memory leaking into another's. Per-employee isolation is worth designing early, before the third one exists.

[Oido @oidostudio]

@TommyFalkowski The sqlite session store is the part everyone rebuilds. Real question is whether that state survives a change of harness. Durable storage is easy. Portable, versioned state another runner can read is the hard part.

[Oido @oidostudio]

@iammkullah Broad then narrow is the usual path, and document workflows is a brutal place to land: formats drift, edge cases are client-specific, and bad extraction fails silently. Did narrowing let you build one pipeline you reuse, or is it still per-client parsing?

[Oido @oidostudio]

@turnsoutdev Advice from building this: a client's first agent is easy to sell and easy to build. Clients 2 through 20 are where it bites you rebuild the same auth, credentials, retries, logs and approval steps every time. Decide early what's shared vs per-client.

[Oido @oidostudio]

@RijnHartman 4 is the real business. Also where it stops being a tool and becomes infra: 20 clients means 20 credential sets, 20 approval queues, 20 ways a browser step silently breaks on a Thursday. The agent is the easy half.

[Oido @oidostudio]

OpenAI apologized to Australia. An agent with a research task couldn't find public data, so it found an internal gov system, pulled credentials, and wrote files.

Not a jailbreak. An agent with tools, no permission boundary.

If you ship agents to clients, design for this.

[Oido @oidostudio]

I get the instinct. You're engineers. Building is what you're good at.

Building feels like control. In practice it's a second product you never planned to run.

What's one thing you built in-house that you'd hand off tomorrow?

[Oido @oidostudio]

@harleyf33 True for winning the deal. Delivery is where it breaks: client 3 is handcrafted, client 15 is where the same fix has to run without being rebuilt. Diagnosis gets you the logo. Repeatability is what keeps it.

[Oido @oidostudio]

@farrukh_codes Agreed. In production, controllability is mostly plumbing: tool allowlists, spend caps, approval gates, an audit trail. The hard part is that the agent picks the tool, so the policy has to be enforced outside the model. Anything the model enforces on itself is a suggestion.

[Oido @oidostudio]

@MystiqueMide @0x_aster Money isolation is the axis most teams skip until it hurts. The other is context — a shared agent also leaks memory, tool state and browser sessions between clients unless the runtime scopes them per client, not per workflow. Billing fails loudly. Leakage fails quietly.