Agent mode

Bootstrap once.
Then watch.

A human signs up, names a project, and mints one key. Claude — or any MCP client — runs the rest: endpoints, fan-out targets, monitoring, incident triage. Two real walkthroughs.

Start free See the demos

Most platforms make you the operator.
Ours makes the agent the operator.

Every action you would have clicked in a dashboard is a tool the agent already holds — 40 of them, over MCP and REST, identical surface either way. The four things only a human can do (sign up, name the project, mint the first key, approve the very first charge in hosted checkout) take under a minute. The rest is hands-off.

Day one — greenfield setup

Stripe fan-out, before lunch.

Stripe webhooks landing in four places — Slack alerts, internal billing DB, an analytics warehouse, and a customer-email queue — without a single line of glue.

You

~60 seconds

  1. 1

    Sign up

    Email + password. No card.

  2. 2

    Create project

    Name it stripe-relay.

  3. 3

    Mint a key

    Paste it into Claude.

Then you close the tab.

Your agent

~3 minutes, hands-off

MCP surface
  1. get_project Confirms the token's project — name, plan, edit access.
  2. list_lines Empty project. Room to build.
  3. create_line Cuts a v1 line for the Stripe surface.
  4. default_endpoint_template Pulls the canonical starter JSON — schema, validators, target shape.
  5. update_endpoint Reshapes the body for Stripe's event payload — type, data.object.id, livemode.
  6. update_endpoint Adds the Slack target — channel webhook URL, condition type == "charge.failed".
  7. update_endpoint Adds the internal billing-DB target — Bearer auth from env, retry policy stays default.
  8. update_endpoint Adds the analytics warehouse — HMAC signature header, longer retry window for batch ingest.
  9. update_endpoint Adds the customer-email queue — filter to customer.subscription.deleted only.
  10. dry_run_endpoint Validates the whole document against a synthetic Stripe payload. No DB write, no billed call. Passes.
  11. create_endpoint Endpoint live at stripe-relay.echorelay.cloud/v1/stripe.
  12. create_api_key Mints the inbound key Stripe will present. Plaintext returned once; the agent stows it in the project's encrypted Stripe-webhook-secret field.
  13. get_config Confirms the live Framework config matches the doc just written. Single source of truth.
  14. list_requests After the first real Stripe event hits — 1 inbound, 4 fan-out attempts, 4 × 200. Clean.
  15. get_metrics Hour-window p50 latency: 87 ms inbound, < 120 ms per target. Posts summary back.

Then it pings you on Slack: "stripe-relay is live. First event landed at 11:47."

Day N — live ops

3 a.m. The agent triages while you sleep.

Per-target metrics + a full request log + a failed-deliveries (DLQ) retry surface — the agent doesn't need you to investigate. You just need to be reachable.

You

A single sentence

  1. 1

    Ask

    "Anything broken?"

Then you go back to sleep.

Your agent

~90 seconds, hands-off

MCP surface
  1. get_metrics Project-wide, last hour. Error rate climbed from 0.4% to 11.2% at 02:51.
  2. list_requests Filter status=502, since=02:51. 86 rows. Same path, four targets per row.
  3. get_request Pulls one for the trace — three targets 200, one 502 with connection refused. The agent has its suspect.
  4. get_metrics Re-runs with targetId on the suspect. 100% error since 02:51. Confirmed degraded.
  5. list_requests Filter targetOutcome=dlq. 23 entries piled up. Nothing's lost — retries already exhausted but the payloads are intact.
  6. list_dlq 23 rows, all on the same target, all connection refused. The agent stages a retry plan.
  7. get_metrics Checks the suspect once more — error rate dropped back to 0.1% at 03:04. The downstream came back.
  8. retry_dlq_entry Re-enqueues 21 of the 23. Each lands a 200 within seconds.
  9. discard_dlq_entry 2 entries are duplicates of events your on-call already handled out-of-band — discarded with a note in the audit log.
  10. list_audit_events Confirms its own work is on the record — actor, action, entity ids, timestamps.
  11. post to Slack Posts the timeline + a one-paragraph postmortem to the on-call channel. Done.

In the morning you read: "analytics.warehouse.internal was down 02:51–03:04. 21 events recovered, 2 deduped. No action needed."

Four things only you can do.

Deliberate. Identity and consent stay human.

  • 1

    Create the account

    Email + password. Once. The agent has no identity to bootstrap from.

  • 2

    Name the project

    One field. Locks the slug your callers will use.

  • 3

    Mint the first key

    Paste it into your MCP client. Subsequent keys the agent can mint, rotate, or revoke itself.

  • 4

    Approve the first charge

    Hosted checkout, one time per account. Subsequent buys the agent charges your saved card directly.

Free tier. No card.
Unlimited requests — rate-limited, not metered.

Try EchoRelay against real traffic. Promote to a paid plan when you outgrow it — your agent can handle that too.

Start free Compare plans
Currency: