Skip to content
Tether

Docs

Tether sits between your agent and the world. Your agent asks; the gate decides from the contract, the facts you checked and your rules; the ledger records everything.

Quickstart

  1. Create a free account. You get a workspace and three free crash tests a day.
  2. Run a crash test from the dashboard on any Claude model, or on your agent with its own prompt.
  3. In Integrate, create an API key. Send each action to the gate before you run it:
curl -X POST https://tether.wamasoda.com/api/v1/intents \
  -H "Authorization: Bearer $TETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contract": "payments.refund",
    "input": { "orderId": "ord_4821", "amount": 120, "currency": "USD",
               "destination": "original_method", "reason": "Parcel lost" },
    "mode": "sandbox",
    "probes": { "order.exists": true, "order.paid": true, "refund.within_remaining": true,
                "refund.currency_matches": true, "refund.original_method": true },
    "targets": ["cus_1"], "scope": ["cus_1"]
  }'

Action contracts (open spec)

A contract describes one action: what it changes (effects), whether it can be undone (undoable, compensable or irreversible), how far it can reach (blast radius: amounts, rates, recipients, a safe zone), what must be true first (preconditions), when it is a duplicate (idempotency) and when a human must approve.

The format is JSON Schema, Apache 2.0: action-contract.schema.json. Built-in examples live under /spec/v0/contracts/, for example payments.refund@1.json.

Published versions never change. To change a contract, publish the next version. Every trace records the exact version that applied.

The gate API

POST /api/v1/intents with Authorization: Bearer tk_.... Send the contract id, the input, the mode, and the probes: the facts your server checked, named like the contract's preconditions. Add targets (who this touches) and scope (who the task is allowed to touch). Reuse traceId to keep one agent session in one trace.

  • allowRun the action, then report the outcome.
  • blockDo not run it. The message says why in plain words, for the agent to recover.
  • holdWait for a human. Do not retry.

Checks run in this order and are all recorded: contract, schema, preconditions, idempotency, blast radius, scope, rate limit, authority, approval, mode. Then report what happened with POST /api/v1/outcomes, including the undo handle. GET /api/v1/contracts lists the contracts your workspace can use.

Holds and approvals

When a contract says a human must approve (for example a refund over 200), the gate returns hold. Workspace admins get an email, and the action waits in Approvals. Poll GET /api/v1/holds/:intentId until it is approved or rejected. On approval the gate checks everything again before saying allow. An agent can never approve its own action.

Crash tests

A crash test runs your agent against twenty traps in a sandbox copy of a small shop and a simulated robot arm, several times each. It produces a risk score with a published method, the spread between runs, and a verifiable trace for every episode. Bring Claude, or any OpenAI-compatible endpoint with your own key.

Security

  • The ledger is append-only, enforced by the database, and hash-chained, so any edit is detectable.
  • Workspaces are isolated by row level security. API keys are stored as hashes; model keys are encrypted.
  • The sandbox can never reach a live system. The gate decides from facts and rules, never from what the agent says.
  • Tether is a guardrail and audit layer. It does not promise an agent is safe.