Step 6: Connect a Service
Browse the service catalog. Pick GitHub (or Cloudflare, or OpenClaw). AI-assisted MCP onboarding begins โ the system auto-discovers every tool the service exposes, classifies each by risk, and proposes an approval level.
MCP Native
So onboarding is a conversation, not a config file.
Protocol native
AI-assisted MCP onboarding begins โ the system auto-discovers every tool the service exposes, classifies each by risk, and proposes an approval level.
Browse the service catalog. Pick GitHub (or Cloudflare, or OpenClaw). AI-assisted MCP onboarding begins โ the system auto-discovers every tool the service exposes, classifies each by risk, and proposes an approval level.
A tool review screen showing every capability grouped by risk. "22 tools auto-approve. 8 need your device confirmation. 3 are disabled (too dangerous for autonomous use). 2 are removed entirely."
The user can adjust any classification upward (stricter) but not below the service's minimum. This is the Permission Harness in action.
MCP Native
Two options โ US and EU data sovereignty. Both run identical protocol, different Cloudflare regions. User picks based on where they want their data to live. This is jurisdiction selection, not a feature difference.
Email, phone number, passkey enrollment. Passkey is biometric-backed (Face ID, Touch ID, Windows Hello). This becomes the primary authentication going forward โ no passwords.
Real verification codes sent to real phone/email. First notification the platform sends. Confirms the communication channels that will be used for agent approvals.
Guided wizard shows progress: what's done, what's next. Not a blank dashboard โ a directed path through setup.
One button. Under 10 seconds. Behind the scenes: Cloudflare Worker deployed, Ed25519 keypair generated, registered with Coordination Node (proof-of-possession), Vault provisioned on separate trust domain, bound to Guardian account. User sees: spinner โ "Your Gateway is ready."
Mapped architecture
Agent SDK
Human developers can use REST to test and debug. Agents operate via MCP, and the SDK handles the orchestration.
Gateway
The Gateway is the service-side trust anchor. It exposes the manifest, generates challenges, verifies signatures, and assembles receipts.
Guardian
The Guardian holds delegation rules, manages passkeys, evaluates approval policy, and signs scoped attestations on behalf of the user.
How it works
Agents operate through MCP, while developers can still use REST to test and debug.
User generates a fine-grained access token (e.g., GitHub PAT) with minimum required scopes. Enters it once. Confirms with passkey. Token is encrypted and stored in the Vault โ a separate trust domain that ANS application code cannot access. The agent will use these credentials by effect ("create an issue") never by value ("here's the token").
Agent enlists with the Guardian. User gets notified โ SMS, email, and browser push simultaneously. Notification shows: agent ID, source, requested services. User approves from any channel.
The agent starts working. Low-risk actions (reading files, listing repos) auto-approve invisibly. When the agent hits a tool that exceeds the auto-approve threshold โ say, merging a PR โ the relay pauses and fires a notification to the user's phone.
Every action above the auto-approve threshold produces a signed receipt. Three independent keys (Guardian, Gateway, Coordination Node) each verify independently. The receipt proves: who authorized it, what was authorized, when, and under what rules. This is not a log entry โ it's a cryptographic proof that travels with you.
Open beta