Both sell the same idea — an AI employee, not a chatbot. The fork in the road is where it runs and who holds the keys. Junior is a cloud service; Hedy deploys inside your walls.
Hedy is a private-deployment AI employee: it runs on your own Docker/K8s or air-gapped infrastructure, uses your own model keys (including local vLLM/Ollama), and prices per seat. Cloud AI employees like Junior run on someone else's servers, meter you by usage credits, and hold your data in their tenant. Hedy keeps source-of-truth, retrieval ACLs, and append-only audit inside your perimeter, and is Feishu/Lark-native alongside Slack. One Hedy replaces a team's output, not a chat seat.
| Hedy | Junior | |
|---|---|---|
| Where your data lives | Your VPC — or fully air-gapped | Vendor cloud (AWS) |
| On-prem deployment | Yes — Compose or Helm | Not offered today |
| Source auditable | Yes, by your security team | No |
| Chat platforms | Feishu & Lark first-class, plus Slack | Slack & Microsoft Teams |
| Model keys | Yours — incl. local models | Vendor-managed |
| Pricing model | Flat per-seat licence | Usage credits, from $100/mo |
| Conduct red-line testing | Shipped in the eval gate | Undisclosed |
| Integration breadth | Core enterprise stack, growing | 3,000+ via connectors |
Based on public information as of July 2026. Corrections welcome: hello@hedy.one
The core split is where the work runs. A cloud AI employee lives in someone else's tenant: your prompts, retrieved documents, and outputs cross their boundary, and you accept their sub-processors. Hedy deploys into your own infrastructure via Docker or Kubernetes, and runs air-gapped when you need it. There are no sub-processors. You bring your own model key and point it at a hosted API or a local vLLM/Ollama endpoint, so inference can stay inside the same network as your data. A typical compose deploy takes about 30 minutes.
Pricing follows the model. Cloud AI employees meter you on usage credits, so cost scales with every token and every retry, and forecasting is guesswork. Hedy is priced per seat, and every LLM call routes through one gateway that owns metering and quota, so spend is one number you control, not a variable bill from a vendor. Licensing is offline: an Ed25519-signed license validates without phoning home, which is what an air-gapped or regulated deployment actually requires.
Governance is the product, not a setting. Retrieval is ACL-scoped to the asker, so an AI employee never surfaces documents a person could not open themselves. Write actions are deny-by-default against an allowlist, with tiered approvals (L0/L1/L2) for anything that leaves a mark. The audit log is append-only with no update or delete path. An evidence gate requires answers to quote their sources verbatim, so claims are traceable rather than plausible. Red-line behavior tests run against the agent instead of trusting instructions in a prompt.
On integration, Hedy treats Feishu/Lark as a first-class channel alongside Slack, so the AI employee works where your team already coordinates rather than forcing a US-centric stack. The framing is different too: a cloud AI employee is one more seat in a chat workspace; one Hedy is meant to produce the output of a team. Private, not cloud. An employee, not an assistant. Doing the job, not recording the meeting.