NapNapParallel threads
Run several task lines at once; each freezes its agent profile and can be pinned, archived or resumed. History lives on your machine.
Shipped
Phase 0 paid beta · zh / en / it
One workspace.
Hand the busywork to your agents.
NapNap Workspace is a desktop AI workspace for e-commerce operators. Parallel threads, configurable Codex agents, Telegram channels, scheduled jobs, skills and MCP — all in one window, with every command and file edit gated by approval.
- 1M context by default
- Read-only / workspace / full-machine
- Threads stored locally
- Credentials in the OS keychain
NapNap Workspace — Inventory check & copy rewrite＋↗◫Workspace
Hui · Team12,480 credits
Inventory check & copy rewriteGPT-5.6 TerraContext 18% / 1M
Check yesterday's Shopify inventory, list what's out of stock, and rewrite the titles.
Working directory connected. Running the inventory sync first, then rewriting titles with the product-copy skill.
commandpnpm run sync:inventoryexit 0 · 1.8s
Sync complete: 214 SKUs · 7 out of stock · 12 low
Approval needed: write file
Write the 7 rewritten titles to ./exports/titles-0917.csv (currently read-only).
AllowDecline
Say something, or call a skill with $…$product-copyAuto ▾↑
Click the rail to switch panels
Capabilities
Not a chat box — a workspace that acts
Each piece maps to a real day of e-commerce work: parallel tasks, explicit permissions, auditable results.
Codex agent workspace
Define identity, skills and memory in md files. Three permission presets, switchable mid-thread from the next message.
Shipped
Skills & MCP
Toggle skills globally or allowlist per agent. MCP supports JSON import, OAuth and bearer tokens.
Shipped
Telegram channel
A six-digit pairing code binds a DM or group to a thread. Customer messages in, replies out, from the workspace.
Shipped
Scheduled jobs
Generate competitor research every Monday morning, pushed into the same thread with a badge when done.
In progress
Knowledge & contacts
Separate knowledge bases mounted per agent; customers and colleagues with roles, recycle bin and audit trail.
Planned
By role
One workspace, four ways to use it
Cross-border sellers
Inventory, copy and customer messages in one thread
Hand off the repetitive listing, repricing and replying; you just confirm.
✓Photo-to-listing: capture on mobile, monitor on desktop
✓Inventory checks and stockout alerts written straight to a sheet
✓Customer questions from Telegram groups turned into a task list
Typical flow
1CaptureScan and shoot on mobile, watch on desktop
2GenerateListing engine writes copy, AI renders the hero image
3ConfirmEdit once; files are written only after approval
4ExportPlatform format out, inventory back to your backend
Why us
Trust comes before features
01
Approval can't be bypassed
Commands and file edits always go through an approval card: idempotent on double-click, no implicit answer on interrupt, no auto-timeout, declines never render as done.
02
Credentials stay off disk
Login credentials are encrypted by the OS keychain, tool shells don't inherit secret variables, and shell snapshots are off.
03
Your threads are yours
Threads and records live on your machine — switch models, restart or drop offline and resume the same context.
04
Bring your own model
GPT-5.5 / 5.6 tiers, GPT-6, five Claude models, Gemini and DeepSeek, switchable inside one thread.
A day's work in three minutes
From photo-to-listing to customer replies, one full pass through the real flow — approval cards and skills included.
▶Video slot (send me the footage and I'll drop it in at this size)
What beta users say
I used to keep three windows open for stock, replies and copy. Now it's one thread — and it always asks before writing a file.
FAQ
Late Phase 0, moving toward paid beta. Chat, agents, skills, MCP and the Telegram channel are usable; scheduled jobs, knowledge bases and contacts are still in progress or planned. Paid plans are activated by contacting [email protected].
GPT-5.5, GPT-5.6 Luna / Terra / Sol, GPT-6 Astra, Claude Fable 5.1 / Opus 5 / Opus 4.8 / Sonnet 5 / Haiku 4.5, Gemini 3.5 Flash / 3.1 Pro / 3 Flash, DeepSeek V4 Pro / Flash. Default is GPT-5.6 Terra, switchable mid-thread.
The default is workspace-scoped. In read-only mode writes require an approval card; full-machine access must be enabled explicitly. Approvals never auto-timeout, and declined actions don't run.
Threads and session state stay on your machine; credentials are encrypted by the OS keychain. Domestic and overseas deployments are planned separately, with mainland users' personal data kept in-country.
The interface ships in Chinese, English and Italian, with Japanese, Korean and Spanish planned. Model replies follow the interface language.
Subscription plus credits: chat billed by token, image generation by run. Until online payment lands, write to [email protected] to be activated.