Challenge
An AI agent with tools can change real records, and it can do something expensive when nobody is watching. A business needs those actions and costs under its control, without giving an outside service access to its database.
Back to projects
AI execution layerInsights
Tether
AI can act. You stay in control.
Tether is a governed execution layer between AI and a business’s systems. An application posts a question and a tenant identity. Tether calls a model with the customer’s own tools, refuses to change anything until the write is approved under the business’s policy, and returns structured JSON plus the exact cost of the turn. It never receives database credentials: it reaches data only through HTTPS endpoints or APIs the customer controls.
Services
AI systems
Solution
Tether sits between the AI and the business’s software. A business connects an OpenAPI document, an MCP server or a signed tool endpoint. It then runs Tether’s assistant on its website, or keeps its own AI and puts Tether in front of its tools. Read tools run during the turn. A write tool does not call the business’s endpoint during the turn: it becomes a pending action, a person approves it or, for low-risk tools, a rule the business set does, and it executes once.
What it does
- Every write is staged as a pending action. After approval it runs exactly once and is never retried, and every transition is audited.
- Spend is checked before the model is called, using an append-only ledger, per-tenant budgets and a spend breaker.
- Tenant scope comes from the API key, never from request bodies or model arguments.
- A business that brings its own AI keeps its model and key, and tool results never pass through Tether to a model provider.
Technologies
- MCP
- OpenAPI
- Next.js
Deliverables
- HTTP API and MCP endpoint
- Embeddable assistant
- Approval inbox
- SDKs for Node, Python and .NET
- Documentation site