EchoRelay MCP Server

Create an endpoint under a line. Defaults to immediate publish; draft mode requires draftRevision. scheduledAt stages the line draft then schedules it. Requires edit access.

Delete an endpoint and its targets, publishing the removal immediately. Requires edit access.

Get the full JSON of one endpoint, including its targets, auth keys and attributes.

Update an endpoint immediately or in its line draft. The patch is merged field-by-field. Requires edit access.

Create a new project on your account so an agent can bootstrap from a fresh account. The slug is derived from the name and validated server-side (format, reserved words, uniqueness). Each plan includes a fixed number of active projects (free tiers one; paid plans more: see list_subscription_plans); at the limit this errors. If an existing project can host this integration, skip create_project and call create_project_token against it instead of adding another. Requires an ACCOUNT-scoped management token and the `config` scope. Returns the created project ({id, slug, name, apiBaseUrl, archived}); call create_project_token next to mint a management token for it.

Mint a PROJECT-scoped management token (`er_mcp_`) for MCP and REST; it cannot authenticate relay traffic. Use it after create_project to configure a fresh project, or for any project you already own. Attenuated by design: the scopes must be a subset of THIS management token's own grant (`read` is always included), expiry is mandatory (1–90 days, default 30, never "never"), and the minted management token (being project-scoped) can never mint management tokens itself. `spend` is human-granted only: no management token, of any scope, can mint one carrying it. Mint a spend-scoped token from the project's panel instead. Requires an ACCOUNT-scoped management token and the `config` scope. Returns the plaintext exactly once; only its hash is stored.

Get the EchoRelay project this token is scoped to: id, slug, name, caller-facing API base URL, whether the token has edit access, and the request-log hot-tier retention window.

List the management tokens scoped to a project you own (any holder): id, label, scopes, createdAt, lastUsedAt, expiresAt, revokedAt: never the secret (only its hash is stored). Requires an ACCOUNT-scoped management token and the `read` scope.

Preview what Save & Publish (or the pending scheduled publish, if one exists) would change: added/removed/changed lines, and within each changed line the added/removed/changed endpoints and targets, plus any project-level setting change. Outbound auth secrets are redacted the same as get_config.

Get the project's published relay configuration — the exact JSON the EchoRelay Framework consumes. Includes pendingPublish with its frozen, redacted config when one is scheduled.

List retained revisions for one line so a prior revision can be selected for rollback.

Save a new reusable credential, sealed with the project's encryption key at write time. Neither this call nor any later read ever returns the secret back — reference it from a target by id (see create_endpoint/update_endpoint's target.credentialVaultEntryId) instead of copying the secret around. `auth.type` selects which fields apply: bearer→token, basic→username+password, apiKeyHeader→headerName+key, apiKeyQuery→paramName+key, publicPrivateKey→secret+key. `destinationHost` is required and permanent: the credential is only ever sent to that host, over https, and no later call can re-aim it.

Delete a saved credential. Blocked (not a permission or not-found error — a validation error) while any target still references it; repoint or remove those targets first.

Rename a credential and/or rotate its secret, publishing the change immediately so every target that uses this credential picks it up right away. Omit a secret field (or the whole `auth` object) to keep the existing value — the vault can never show a secret back to confirm it, so leaving it out means "unchanged," not "cleared." The destination host cannot be changed here: create a new credential for a new destination.

Permanently delete a DLQ entry without retrying. Requires edit access.

Discard one endpoint change from a line draft.

Discard every unpublished change and schedule for one line.

Create a new versioned line. A new line has no endpoints, so creating one publishes nothing and serves no traffic yet — add an endpoint, which publishes the line with it. Draft mode requires draftRevision 0. Requires edit access.

Delete a line and every endpoint under it, publishing the removal immediately. Requires edit access.

Read one line draft, its optimistic revision, preview, and schedule.

Mint an inbound relay (data-plane) API key for callers to send traffic to this project. Its `er_live_` or `er_test_` secret authenticates relay requests; it cannot call management MCP or REST. Returns the plaintext exactly once — store it now, it cannot be retrieved later. Editor + owner; subject to the project key policy (see get_key_policy).

List the project's inbound API keys (metadata only — the secret plaintext is never returned).

Cancel one line schedule before its two-minute lock.

Schedule the project owner's subscription for cancellation at the end of the current billing period. Returns `{kind: 'scheduled', newPlan: 'free', effectiveAt}`. Add-ons keep renewing while the plan is active; after it ends they stop renewing and each already-paid add-on cycle stays active until its own end date (use unsubscribe_addon to stop one sooner). Owner-only; requires a token minted with the billing scope.

Get DLQ entries by entry id (`{ms}-{seq}`, as returned by list_dlq's `id`) or by `requestId`. `requestId` is the durable handle — stable across a retry, unlike `id`, which changes every time an entry is replayed and later dead-letters again — and returns every fanned-out target's entry for that inbound request (one request can fan out to N targets, and several may dead-letter); `id` returns at most one. Each entry includes configVersion (the published config that authorised the delivery; 0 means unstamped). Errors if nothing matches this project — purged, retried, or discarded entries age out the same as any other.

List entries currently in this project's dead-letter queue, newest first. One inbound request fans out per-target, so a single request may produce several DLQ entries, one per target that didn't accept it (different `targetId`s). Returns `{total, limit, offset, rows[], evictedCount}` where each row has `id` (the dead-letter entry id, `{ms}-{seq}`), `requestId`, `targetId`, `configVersion` (the published config that authorised the delivery; 0 means unstamped), `failureReason`, `failureCode` (stable machine-readable cause; `unclassified` for entries written before codes existed), `failedAttempts`, `failedAt`, `payload` (the JSON of the queued delivery as it was attempted); `evictedCount` is the lifetime count of entries the queue's capacity cap discarded before they could be triaged. DLQ entries — including the original request body and headers — are kept for up to 30 days from the failure time or until cleared, then purged automatically (or discarded early past capacity — see `evictedCount`); they are never written to a database. get_request still answers what happened to a purged/evicted/discarded entry's inbound request for the project's requestLogRetentionDays window, independent of whether the DLQ row itself still exists. `requestId` is the durable handle across a retry: an entry's own `id` changes every time it is replayed and later dead-letters again, so get_dlq_entry accepts a `requestId` lookup as well as `id`.

Returns the project's key policy: defaultKeyTtlDays, maxKeyTtlDays, editorsMayCreateNonExpiringKeys.

Owner-only. Edit the project's key policy. Only fields present in the patch are touched. Editors calling this get an error — use get_key_policy to read.

Remove a member (accepted or pending) from this project by their member ID or email. Owner only.

Change a member's role between editor, viewer, and billing. Identify the member by their member ID or email (from list_members); the project owner's own role cannot be changed. Returns the updated member record. Owner only.

Preview one line draft without publishing it.

Preview what would happen if change_plan were called with this plan: the immediate-charge amount (cents), the next-invoice amount, the effective date, and the kind (`applied` for in-cycle upgrade, `scheduled` for at-period-end downgrade / cancel). Owner-only; requires a token minted with the billing scope.

Manual hard revoke. The key stops authenticating on the next request. Owner anywhere, editor on own keys.

Revoke a PROJECT-scoped management token on a project you own, any holder's, not just your own (the account-owner kill switch for a leaked or retired integration's management token). Soft-revoke: the row stays for traceability and the management token is rejected immediately. Idempotent. Requires an ACCOUNT-scoped management token and the `config` scope.

Begin a new Pro or Scale subscription for the project owner. Returns `{checkoutUrl, plan}` — surface the URL to the human for approval. Errors if the owner is already on an active paid plan (use change_plan instead). Owner-only; requires a token minted with the billing scope.

Begin a credit-pack top-up for the project owner. Pass the pack`s `eurCents` (from list_credit_packs). Returns `{checkoutUrl, credits, eurCents, directCharge}` or `{transactionId, credits, eurCents, directCharge:true}` on subsequent MCP purchases. Owner-only; requires a token minted with the billing scope.

Get the project owner's current subscription: plan, active flag, period start/end, cancel-at-period-end flag, external Paddle subscription id, and whether a payment method is on file at the provider. Owner-only; requires a token minted with the billing scope.

List the available subscription plans (Free, Pro, Scale) with monthly EUR price, credit allowance, per-project RPM cap, and how many active projects the plan includes. Read-only.

Activate and immediately publish an inactive line so it serves traffic again. Blocked at the active-line cap — deactivate another line first. Requires edit access.

List the project owner`s currently-active add-ons. Returns `[{addonKey, project, monthlyEurCents, activatedAt, nextRenewalAt, consecutiveFailures}, ...]`. RPM upgrades are per-project; `project` is the slug they apply to (null for account-wide add-ons). Owner-only; requires a token minted with the billing scope.

List the optional add-on subscriptions (RPM upgrades, extra team seats, extended log retention, extra file-delivery storage in +5 GB blocks, …). Each entry is monthly, renews every 30 days until cancelled, and carries `scope` (per_account or per_project). Read-only.

Archive a project you own. Reversible: the project + slug persist and credits are kept, but the data plane drops the tenant (callers get 404). Idempotent. Fails with an error if the project has archive protection enabled. That is a support-only unlock, not something this call can override. Requires an ACCOUNT-scoped management token and the `config` scope.

List project audit-log entries, newest first. Captures who changed what — lines, endpoints, targets, API keys. Outbound-target `auth.token` / `auth.password` are redacted in the diff per the same policy used for endpoint reads. Retention is the `auditRetentionDays` advertised on get_project (default 365 days); rows older than that are purged by the cleanup job. Returns `{total, limit, offset, retentionDays, rows[]}` where each row has `id`, `createdAt`, `actor` (email or null), `action`, `entityType`, `entityId`, `entityLabel`, `diff`.

Get the project's credit balances (paid + testing) and which ledger it settles on. `billingModel` is `prepaid` (metered against the credit balance) or `invoiceable` (an Enterprise account, settled on the contract — its credit fields read empty/zero by design, not because it is out of credits). A zero balance is not on its own a reason to stop: live traffic is never refused for an empty balance, and a free-tier project is exempt from live credit billing entirely. Only the testing pool refuses when exhausted (402 `testing_credits_exhausted`). Read this before concluding from any other field that a project cannot send.

Enable or disable the auto-decaying burst window for live traffic and publish the change immediately. When enabled, brief spikes a little above your steady RPM cap are absorbed at the normal credit rate (no surcharge, no burst SKU) instead of being rejected. Sustained overage decays back to your steady cap; headroom returns once traffic cools below the cap or you add RPM. Burst usage is tracked so you can see when to upgrade. Default is off (steady cap is a hard 429). Requires edit access.

Switch the project owner between Pro and Scale, or cancel the subscription (effectively dropping to Free at period end). Returns `{kind, newPlan, effectiveAt, applied}` — `applied=true` means Paddle was charged immediately, otherwise the change is queued for the current period end. Use preview_plan_change first to see the dollar impact. Owner-only; requires a token minted with the billing scope.

List the project's saved credentials — reusable target-auth secrets that can be referenced from any number of targets instead of entering auth inline per target. `auth` is redacted (secret fields never returned, even encrypted — see create_credential for why).

List the prepaid credit packs the project owner can purchase. Read-only. Identify a pack to start_topup by its `eurCents`.

Take a line off the data plane immediately (callers get 404) without deleting it — endpoints/targets/config are preserved and it can be reactivated later. A deliberate action; live traffic is never auto-disabled. Requires edit access.

Return a canonical minimal-valid endpoint JSON for a fresh line. POST it verbatim to create_endpoint to land an immediately-working endpoint, then customise. It comes back with `mock: true`: it answers callers with its own canned body and never dials the target it carries, so a 200 from it proves the endpoint is reachable and nothing about delivery. Set `mock: false` and give the target a real URL before treating a call as delivered. Read-only.

Turn the line's public OpenAPI docs on or off. When enabled, the line's spec is published at docs.{baseDomain}/{slug}/{lineKey} (and …/openapi.json); when disabled those URLs return 404. The `docsShared` field on each line read (list_lines) reflects the current state. Requires edit access.

Schedule a tier downgrade for a non-stackable add-on (RPM tier). The current tier stays active until `nextRenewalAt`; at that point the lower tier is charged and activated. No mid-cycle charge. Use list_addons to find available keys. Returns {addonKey, targetAddonKey, project, effectiveAt}. Owner-only; requires a token minted with the billing scope.

Validate an endpoint document without writing it. Returns either {ok: true, resolved: <Framework config slice>} or {ok: false, errors: {...}}. Pass endpointId to dry-run a PATCH against an existing endpoint; omit it to dry-run a create. Supply sample and/or answerSample to also run each target's mapping against a real payload and see what it produces, which field came from which node, and every rule that failed with the value that broke it. No DB write, no queue entry, no billing, and the payload is never sent anywhere.

Copy a line and all its endpoints/targets/rules into a new line key — build v2 from v1. The copy is saved inactive and remains off the data plane; edit it, then activate_line when ready. Requires edit access.

List the endpoints under one line of the project.

Generate a draft endpoint under a line from a sample inbound request and a description of the target: infers the inbound body schema from the sample, reads the target's method/URL/fields from an OpenAPI document (targetDocument) or a bare sample plus targetUrl, and proposes an outgoingMapping pairing every target field to a same-named inbound field. Nothing is guessed — a target field with no same-named match is reported in the "unmapped" list instead of being mapped, and every field it did map carries "inferred": true in the stored mapping tree. Always writes a draft (never publishes); the draft flows through preview_line_draft, dry_run_endpoint, config_diff and publish_line_draft like any other. Requires edit access.

Invite a person by email to collaborate on this project. Returns the new invitation record; the invitee receives an email with an accept link. Enforces the seat cap for the plan tier. Owner only.

List the project's versioned API lines (e.g. v1, v2), each with its endpoint count.

List accepted and pending members of this project. Shows name/email, role, and whether the invite has been accepted. Owner only.

Aggregate request metrics for this project over a time window — volume, error rate and p50/p95/p99 latency, bucketed for charting. By default the series cover inbound requests; pass `targetId` to slice over one outbound target's per-attempt rows instead (answers "which target is degraded?"); pass `endpointId` to slice over inbound rows for a single endpoint (answers "is this endpoint receiving traffic / erroring?"). Pass at most one of targetId / endpointId. Free plans see "hour" and "day"; "week" and "month" require an active paid subscription on the project owner and are silently downgraded to "day" otherwise (the response includes `planGated: true` when that happens). Returns `{window, requestedWindow, planGated, scope, targetId?, endpointId?, bucketSeconds, buckets[], volume[], errors[], totalVolume, totalErrors, errorRate}` plus latency series whose keys depend on scope: an inbound read (project-wide, or `endpointId`) carries `relayP50[]/relayP95[]/relayP99[]` for our own time and `targetP50[]/targetP95[]/targetP99[]` for the destination's, never blended; a `targetId` read carries `p50[]/p95[]/p99[]` for that target's attempts — series arrays are dense (one point per bucket, zero-filled when no traffic).

Narrow the project's outbound host allowlist — the hosts a target may point at. The current list is on get_project as `outboundHostAllowlist` (empty means any public host). This tool may only narrow it: `patterns` must be a strict subset of, or equal to, the current list, compared as exact strings — submitting a wildcard that would cover an existing exact host, or clearing to empty, is refused as a widen. Widening the list is web-only, in the panel. Refused when the project already has a target (inactive lines included) outside the submitted list. Requires edit access.

List the projects on your account (archived included), each with `id`, `slug`, `name`, `apiBaseUrl` and `archived`. Requires an ACCOUNT-scoped management token (one minted with no project) and the `read` scope. A PROJECT-scoped management token cannot call this: use it on its own project's tools instead.

Turn on archive protection for a project you own: archive_project (on every surface: panel, REST, and this tool) then refuses until it is lifted. Idempotent. Permanent from here: there is no tool or API call to turn it back off; only contacting support can. Requires an ACCOUNT-scoped management token and the `config` scope.

Publish one line draft now and clear only its schedule.

Get every target this request fanned out to, each with its delivery receipt or the reason it has none. A still-moving delivery (queued/retrying) answers from the live tracking copy, a terminal one (delivered/failed) from durable storage — the response shape is identical either way, so poll this on an interval with the same requestId a 202 response returned and watch `status` move to a terminal value. Each entry has targetId, targetURL, available, and when available: status (queued/retrying/delivered/failed), attempt, enqueuedAt/updatedAt/completedAt (unix ms), terminalError, callbackState (none/pending/delivered/failed — whether the target's completion-callback URL, if any, has been notified). When available is false, `reason` is "sync_or_stream_delivery" (the target answers inside the original HTTP response and is never tracked here — see the response body from that call instead) or "not_found" (no record for this target, e.g. aged past the retention window).

Rename a project you own (the display name only: the slug, which identifies the project in URLs and API paths, never changes). Requires an ACCOUNT-scoped management token and the `config` scope. 404-equivalent error when you do not own the slug.

Get one inbound request envelope + every per-target delivery attempt made on the way out (success / permanent_failure / retry_scheduled / dlq). The request and each attempt carry configVersion, the published config that authorised them; 0 means unstamped. Scoped to this project; returns null once the request has aged out of the project's log-retention window (requestLogRetentionDays — see get_project).

List inbound API requests the relay has processed for this project, newest first. Hot-tier window is the project's requestLogRetentionDays (see get_project). Optional filters narrow the result. Returns {total, limit, offset, rows[]} where each row has ts (unix ms), requestId, method, path, status, durationMs, targets, creditsDeducted, authType, errorMessage, bytesSent, configVersion (the published config that authorised the request; 0 means unstamped), and closeReason (for a streaming request: complete / abortedOrError / byteCap / noValue; empty for non-stream requests).

Re-send the invitation email for a still-pending invite, by member ID or email (from list_members). The original accept link is reused. Already-accepted members are rejected. Owner only.

Resume (un-cancel) a pending-cancel add-on whose paid-through cycle has not yet ended. Clears the cancellation so the add-on renews normally at the next cycle boundary. No new charge — the cycle was already paid. Stackable add-ons cannot be resumed; buy a new unit instead. Returns {addonKey, project, nextRenewalAt, resumed}. Owner-only; requires a token minted with the billing scope.

Re-enqueue a DLQ entry to the main relay stream and delete it from the DLQ. A replay is billed like any relay — it charges the entry's original credit cost to the project before re-sending; if the balance is too low the entry stays in the DLQ and this returns an insufficient-credits error (top up and retry). Succeeds if the entry exists, errors if it's already gone (retried or discarded). Requires edit access.

One-shot retrieval of the plaintext for a successor minted by the auto-rotation scheduler (not by rotate_api_key — that returns its own plaintext directly). Decrypts the at-rest ciphertext, returns it once, then destroys it; a second call errors. Only available while the predecessor still has an unrevealed successor.

Preview or apply publishing a retained line revision over what is currently published, clearing any pending schedule. Without confirmRevision this only returns the diff (the redacted change-gate preview, same shape as config_diff) between what is currently published and the target revision — it changes nothing. To apply, call again with confirmRevision set to the diff's currentRevision; if a publish landed since the preview, confirmRevision is now stale and the call is refused with error "rollback_confirmation_conflict" instead of silently overwriting the newer publish — read the fresh diff and confirm again. Applying also requires confirmUnsigned, the preview's unsignedFingerprint, so a signing secret expiring between the preview and the confirm cannot roll a target back unsigned without you having seen it named.

Mint a linked successor for an existing key. Returns the successor plaintext exactly once — store it now. The predecessor stays valid through its overlap window so callers can swap without downtime. Owner anywhere, editor on own keys.

Set or clear a self-set ceiling on this project's own effective live-traffic RPM (Traffic shaping) — for when your origin can't take the throughput your plan grants. Never plan-gated: the ceiling can only LOWER your effective RPM below your plan/add-on rate, never raise it, and never touches billing. Steady traffic is capped at the ceiling; if Burst Protect is on, its headroom is measured above the ceiling, not above the plan rate. The current value is on get_project as `rpmCeiling` (null = no ceiling). Requires edit access.

Schedule one non-empty line draft. Active subscribers required.

Subscribe the project owner to a monthly add-on (RPM upgrade, team seat, extended log retention, extra file storage in +5 GB blocks, …). Charges the saved payment method immediately, then auto-renews every 30 days until `unsubscribe_addon` — or until the subscription ends, after which the already-paid cycle stays active to its own end date and stops without further charges. Requires an active Pro/Scale subscription and a saved PM. RPM upgrades (rpm_*) are per-project and apply to THIS project automatically. Returns `{addonKey, project, monthlyEurCents, transactionId, activatedAt, nextRenewalAt}`. Owner-only; requires a token minted with the billing scope.

Restore an archived project you own to the data plane. Restoring consumes one of your plan's active-project slots, so at the limit this returns an error telling you to archive another project or upgrade. Idempotent for an already-active project. Requires an ACCOUNT-scoped management token and the `config` scope.

Cancel an active add-on subscription. Stops future charges; the current cycle stays active until expiry. RPM upgrades are cancelled for THIS project. Returns `{addonKey, project, cancelledAt}`. Owner-only; requires a token minted with the billing scope.