What your twins can and cannot do
This matrix exists so you never promise something the platform will not do. Every row points at the code that enforces it; anything unimplemented is labelled Planned.
Can do
Read your intent in plain language and turn it into a strict intent signal (JSON)
Live todaysrc/lib/sip.server.ts
Delegate work from the Master Twin to a Specialist Twin
Live todaysrc/lib/delegation.server.ts
Search its own tool catalogue and invoke only tools inside its scope
Sandbox by 31 Aug 2026src/lib/tool-registry.ts
Execute through a connected adapter: Gmail, Telegram, signed webhook, WeChat, BSC
Live todaysrc/lib/execution.server.ts
Draft an artifact for you to review and send yourself
Live todaysrc/lib/execution-capability.ts
Pause on an evidence-rich Decision Card before anything consequential
Live todaysrc/lib/refund-evidence.server.ts
Sign every outcome — including a rejection — and anchor it to BSC Testnet
Live todaysrc/lib/receipt-signer.server.ts
Record a prompt trace of what the model could see
Sandbox by 31 Aug 2026src/routes/api/chat.ts
Run entirely on a self-hosted GPU box with a local model
Live todaysrc/lib/ai-gateway.server.ts
Cannot do (refuses by design)
Execute anything when no live adapter can perform the action
The run halts as blocked / no_live_adapter and the output is labelled "draft — not executed". AiXin refuses to simulate success.
src/lib/execution-capability.ts
Move money without an approved Decision Card
Money-moving SIP actions are tiered high risk and require human approval before execution.
src/lib/sip.server.ts
Use a skill it was not assigned
Scope is resolved before invocation; out-of-scope calls are rejected, not negotiated.
src/lib/tool-registry.ts
Run forever, or spend without a ceiling
Step, token, wall-clock and monthly budgets are checked pre-flight, plus an operator kill switch.
src/lib/run-budget.ts
Browse the open web at will
A twin sees only the data sources and adapters you connected. There is no general web-browsing tool.
src/lib/tool-registry.ts
Replay a past run deterministically
Prompt traces are recorded, but replay is planned — never demo it as working.
planned (W6)
Sell itself inside the app
Paid marketplace commerce, entitlements and payouts are planned (W5). Invoice outside the platform for now.
planned (W5)
How this differs from other agent systems
| Dimension | AiXin | OpenClaw / Hermes / dsh |
|---|---|---|
| What it optimises for | Evidence a sceptical third party accepts | Capability and runtime composability |
| Approval | The product: fail-secure, evidence-rich, signed either way | A plugin or a policy setting |
| When no tool can do the job | Halts as blocked and labels the output a draft | Usually returns confident prose |
| Audit trail | Signed receipts, anchored, publicly verifiable | In-process logs |
| Relationship | A trust layer that can run on top of a harness | The harness itself |