1/5 🚀 PilotDeck is now live — the open-source AI agent OS built for all scenarios！ Built by TsinghuaNLP × ModelBest × OpenBMB × AI9stars, PilotDeck is here for full memory transparency, intelligent cost routing, and agents that never stop working for you. 🔗 Show Case: 💻 GitHub： One... person. One fleet of agents. Ship something real. 🔥🧵show more

2/5 WorkSpace: One Project, One Cockpit Most agents give you a chat window. PilotDeck gives you a WorkSpace. One WorkSpace = one complete project cockpit: 🔷 Files — a dedicated filesystem, AI-generated outputs clearly labeled 🔷 Memory — project goals, progress, constraints, and your personal preferences, all stored per-project 🔷 Skills — install from the Skill Store No context bleed between projects. No rebuilding context every session. Just pick up where you left off — or let your agents do it while you're gone.

3/5 📉 Smart Routing = Massive Cost Savings. Why pay top model prices for simple tasks? PilotDeck automatically detects difficulty and routes: ✅ Hard tasks → powerful models ✅ Easy tasks → cheap & fast models Real results :👇

4/5 🧠 Memory: White-box & Project-Scoped Ever had an AI quietly remember the wrong thing — and you couldn't figure out where it went wrong? PilotDeck's memory is fully White-box：every read and every write is traceable end-to-end. And it's focused on what actually matters: 🔷 Project Memory — the goal you're working toward, how far you've gotten, and the boundaries the agent needs to respect 🔷 Feedback Memory — your preferences, your taste, your requirements, refined every time you push back or correct it 🔷 User Profile — Your personal profile, shared across projects with no need to rebuild. No noise. Just what the agent actually needs to do the job right.

5/5 ⚡Always-on: Your Agents Don't Clock Out Most agents wait for you to type something. PilotDeck's agents are already working. Always-on means your agents: 🔷 Proactively find tasks worth doing — without you asking 🔷 Run in isolation / Executed in an isolated environment 🔷 Confirm with you when a decision needs a human call 🔷 Report back on progress, unprompted 🔷 Deliver outputs as actual local files, not just chat bubbles Step away from your desk. Your fleet keeps moving.

@TsinghuaNLP wait. same performance as Sonnet 4.6. one sixth the cost. I had to read that twice

@TsinghuaNLP Impressive direction for transparent production ready AI agents

@TsinghuaNLP 1/6 of the cost matching Claude Sonnet 4.6 performance? That's a bold claim.

@TsinghuaNLP me at 11pm: deploys 6 WorkSpaces me at 11:01pm:

@TsinghuaNLP @Presidentlin

@TsinghuaNLP Incredible！

@TsinghuaNLP okay I've restarted the same agent project from scratch THREE times this month because context kept getting polluted. just found out PilotDeck has per-project WorkSpace isolation and I'm genuinely a little emotional

@TsinghuaNLP WorkSpace isolation has you covered🥰

@TsinghuaNLP "One person commanding an army of AI agents" is either thrilling or terrifying, possibly both tbh.

@TsinghuaNLP Claude Code: opens a folder Cursor: opens a folder PilotDeck: opens a cockpit

@TsinghuaNLP Being able to manually edit or delete what the agent memorized is a nice touch. I'm so sick of clearing the entire chat history just because an agent hallucinated on step 3 and ruined the context. Looking forward to testing this.

@TsinghuaNLP Been waiting for someone to kill the "one chat = one project" assumption in agent tooling PilotDeck's WorkSpace is what I actually want — isolated memory, isolated files, isolated skills, per project and it's open source！

@TsinghuaNLP Come inbox 📥

One person, one fleet of agents is exactly the direction this needs to go—memory + cost routing is where “agent OS” becomes real infrastructure rather than demos. Curious how the routing behaves under mixed workloads (long memory vs fast tools) and how transparent the cost decisions are in practice.

@TsinghuaNLP Open-source agent OS is where everything is heading. The layer most of them skip is the one that makes the agent actually yours, your memory, your identity, your rules traveling with it. Infra is getting solved. Ownership isn't.

@TsinghuaNLP @grok What are the advantages of this project over Claude Code?

@TsinghuaNLP Great project. Follow back or dm me 📩🤝

@TsinghuaNLP Every other tool calling a folder a "workspace" 👉👈 PilotDeck workspace with memory, skills, isolation, and cost tracking

If you are building multi-step AI agents, context pollution is your worst nightmare. PilotDeck solves this by creating a true "white-box" sandbox for each project. You can literally track, trace, and edit what the AI remembers. This is how real productivity tools should look like.

@TsinghuaNLP Solid work from the team 🔥 Let's connect and build together.

@TsinghuaNLP Interesting approach on the workspace layer. Most AI tools just dump everything into a single prompt session, which breaks down the moment you start multitasking. If the memory isolation here actually works as claimed, it solves a huge headache for parallel projects.

@TsinghuaNLP step 1: set up a WorkSpace and go to sleep step 2: agent finds tasks, executes, files output step 3: wake up to finished work 😌

@TsinghuaNLP turning on intelligent routing and watching your token bill drop 70%

@TsinghuaNLP Exciting stuff! Love seeing open-source innovations like this. Can't wait to see what people create with PilotDeck!

Other AI tools watching your context window hit the limit and forgetting everything. PilotDeck: memory isolated, indexed, editable, rollback ready 🧘

Honestly the most underrated problem in multi-agent workflows is memory contamination across projects. it's not sexy to talk about but it quietly breaks everything. WorkSpace isolation isn't a nice-to-have, it's a prerequisite for anything serious. glad someone finally built it right

Pam holding: "paying $18 for a worse benchmark score" and "paying $3 for a better one"

@TsinghuaNLP ❌ Closing laptop and hoping the AI remembered what you were doing ✅ PilotDeck WorkSpace: still running, still filing, will report back

@TsinghuaNLP People who hate re-explaining preferences to AI every session. People who want their token bill to make sense. both: PilotDeck.

@TsinghuaNLP Great share

@TsinghuaNLP amazing！

@TsinghuaNLP This is basically going from chatbots to full operating systems. Scoped memory + smart routing + cost control is a serious shift in how agents get deployed. 🚀

@TsinghuaNLP Wow, that’s amazing!

@TsinghuaNLP Always-on is not a cron job. The agent actively decides what's worth doing. That distinction matters.

@TsinghuaNLP Routing complex tasks to capable models and simple tasks to cheaper ones is such an obvious idea in hindsight. surprised nobody shipped it properly before this!