Overview
Connect a Gmail support account to a shared Front inbox, poll automation-assigned email conversations and send reviewed Voiceflow answers through Front’s Core API.
Front also supports Microsoft 365 and SMTP-forwarded email channels, plus its External Agents features. This completed recipe chooses a full Gmail account and the Core API; a Gmail alias without its connected primary channel can be outbound-only. Front’s alias guide explains that limit.
Before you begin
Prepare the agent, platform administrator and relay operator.
Use an existing support agent with approved public Knowledge. The agent owner publishes the intended Environment; the help desk administrator owns email routing and assignment; a backend developer operates the relay.
This supervised example answers public support questions. Email addresses and help desk records do not authorize access to private customer data. Keep AUTO_SEND=false: APIs do not consistently expose raw automatic-response headers, so a person reviews every outbound message and its recipients.
Prepare the relay
Install the supplied implementation and durable review queue.
Configure the executable service
Extract the relay bundle, enter email-relay and run with Node 24.14.1:
Use the generated value for IDENTITY_SECRET. Set VF_PROJECT_ID, VF_API_KEY, VF_ENVIRONMENT, a fixed UTC ACCEPT_AFTER just before the pilot and ALLOW_SENDERS to your test inbox. Keep AUTO_SEND=false and POLL_SECONDS=60. The provider steps below supply the remaining fields.
The bundle’s helpdesk-adapters.mjs implements the exact provider calls; helpdesk.mjs checks ownership and input, and the shared worker stores message identities, sessions and cached replies. Read its HELPDESKS.md for file/operation mapping.
Keep one worker and a narrow pilot
Use one process/database for this provider, address, automation owner and scope. The relay scans the selected pilot work and paginates histories; it does not use a webhook or expiring event subscription. API/OAuth credentials still require renewal or rotation. Its 300-page cap fails visibly rather than silently truncating a large scan.
This is a small supervised deployment with no distributed lease or bounded work queue. Monitor quota, disk and oldest pending work; add incremental ingestion and capacity controls before a high-volume rollout.
Connect receiving and replies
Configure the native email channel and its API permissions.
Connect the Gmail account to Front
In Front workspace settings, create the shared inbox, then choose Channels → Connect channel → Google → Gmail Account. Sign in to the support account and grant Front the permissions it requests. Choose whether to import history; keep the relay cutoff after historical mail. Confirm that a new external email arrives and a manual Front reply is delivered. Follow Front’s Gmail setup.
Assign the pilot’s new email conversations to a dedicated automation teammate. Keep the inbox/channel used by this pilot separate from other automation. Existing imported archived conversations are outside its open assigned work.
Authorize the Core API and copy native IDs
Create the token under Settings → Developers → API Tokens, with read access for inboxes, channels, conversations/messages and messages:send. Set PROVIDER=front and HD_TOKEN to that token. Front resource IDs include their prefixes; retain them.
From a pilot conversation’s API response, set HD_OWNER_ID=tea_… from assignee.id. Read GET /conversations/{id}/inboxes for HD_SCOPE_ID=inb_…, and the inbox’s channels for FRONT_CHANNEL_ID=cha_…. Verify that channel is the connected Gmail support account and its address matches all three mailbox/sender fields. The token must have access to that inbox and permission to send for that teammate.
Poll inbound email and send on the original conversation
The relay paginates the inbox’s conversations and each conversation’s inbox memberships/messages. It requires the configured assignee, an open assigned conversation and an inbound email addressed to MAILBOX. Drafts and outbound messages cannot start a turn.
It uses Create message reply, POST /conversations/{id}/messages, with the configured email channel_id, automation author_id, one customer recipient and options.archive=false. A comment endpoint creates internal discussion; this reply endpoint sends email.
To inspect the native IDs, first fill this provider’s credentials, then run the supplied read-only helper in a private terminal:
Replace any PILOT_…_ID with the native ID from a ticket/conversation you can open. The helper permits ID discovery before owner/scope fields are filled; it only performs GET requests plus any required OAuth token exchange. Its output can contain customer data. Copy the fields identified above, then run the scoped check below.
The supplied adapter builds this reply payload after approval; replaceable IDs come from the native incoming record, not from the language model:
Run the persistent worker
Verify read access, then start processing without a public webhook.
Check the native route without sending
After filling .env, run:
Expect your test ticket/conversation with the correct owner, open=true, scoped=true and eligible=true. This command may obtain/cache an OAuth token but sends no Voiceflow turns or customer replies. Resolve access/routing problems before starting the worker.
Start locally, then host the same state
In another terminal, curl -f http://127.0.0.1:8080/health should report running. No public /inbound endpoint is enabled for help desks. Polling uses outbound HTTPS and stored server credentials; there is no incoming webhook secret or subscription to configure.
For a persistent Linux host, install Docker/Compose, copy the bundle and protected .env, set DB_PATH=/app/data/relay.sqlite, and run docker compose up -d --build. The supplied service has a persistent volume and restart policy; port 8080 is bound to host loopback. Keep it private. If you obtained a local OAuth grant, stop the local worker and copy its complete stopped database into the host volume before starting; never run both copies. The bundle README gives the exact database copy commands.
Keep tickets and sessions aligned
Use native ownership, message identity and the actual customer sender.
Use Front’s conversation ID, not its optional human-facing ticket number, as the native session key. The sender address isolates customers within it. Front owns the Gmail channel’s reply/thread behavior; the relay never sends a parallel Gmail API response.
Reassign to a person before replying manually. The recheck is not a universal lock: keep competing humans/automations outside this assignment. Resolve old held jobs before returning work to automation.
For additions during review, show identifies helpdesk.priorJobs and the hold reason. Inspect those cached answers and all new questions, discard stale work, then write one reconciled reply for the newest held job; the agent does not replay earlier messages.
Files stay in the help desk and are never fetched or forwarded. Check original quoted/inline content; the quote heuristic is limited. Outgoing/agent messages cannot start self-reply loops. Manual review also screens external auto-responses whose raw headers the API omits.
Test the complete conversation
Prove receiving, agent execution and actual email delivery separately.
Receive, review and deliver the first answer
From the permitted external inbox, email the support address. Confirm the native ticket/conversation and its automation assignment, then inspect the queue and Voiceflow transcript:
Replace JOB_ID with the actual ID from list. Approve only after checking the question, cached answer, original sender/recipients and support identity; the next worker pass sends it. For review-input, have a person inspect the original and write a reply with the shared review workflow. Expected result: one runtime answer, one reviewed native reply and an email in the test customer’s inbox.
Check the outbound message, channel and customer inbox. Front’s accepted message response is not a delivery receipt. Inspect drafts/error types and the connected Gmail account if delivery stops.
Test continuation, takeover and recovery
Reply in-thread with a follow-up requiring earlier context. Expect the same native conversation and sender-scoped runtime session. Send a separate new support request and expect a new session. Repeat polling and restart the worker; expect no duplicate answer for an already processed native message.
While a reply awaits approval, reassign the ticket to a person, then approve the cached job in this controlled test. Expect ignored and no email. Also test a human reply, two consecutive customer messages, an attachment, an auto-response and a changed Reply-To where supported. Reconcile/discard held work before assigning the conversation back. Finally exercise a controlled API rejection and interrupted send, then use recovery below.
The email never becomes eligible work
Check the native email channel before the agent.
Confirm a real incoming email record, open status, exact automation owner, scope/address and pilot cutoff. The read-only check reports these separately. Check sender allowlisting and provider API permissions; a healthy process cannot prove the support mailbox is receiving. A private note, outgoing email, archived ticket or unrelated channel should not run the agent.
The answer is held or becomes stale
Inspect the durable state and original conversation.
review-input needs an operator-written answer after inspecting the original; review has a cached candidate. ignored with a ticket-change explanation means the ownership/content guard prevented a stale reply. Do not force it through by resetting the database. Discard old held work after the human takes responsibility. Check the released Environment and runtime transcript for uncertain-turn.
The reply was accepted but no email arrived
Reconcile provider delivery before any retry.
Check token scopes, inbox access, teammate API permissions and the channel ID. Merged conversations can redirect; the relay holds instead of following that redirect. Inspect the canonical conversation and resolve old queued work before enrolling it again.
Keep a timed-out write in uncertain-send until native history proves whether it was accepted. Never rerun the agent to recreate a cached answer. A 429 backs off using Retry-After; the sample never retries a write automatically.
Release and maintain the connection
Keep review, assignment and durable recovery in place.
Release the reviewed support route
Publish the reviewed Voiceflow Environment, deploy one worker and repeat the entire customer test through the production support address. Keep AUTO_SEND=false; the example enforces it for help desks. Widen ALLOW_SENDERS only with approved public-support scope, abuse controls and staffed review. Record API/seat access, email delivery and takeover acceptance in your deployment record.
Changing provider, address, owner/scope, project, Environment, cutoff or identity secret requires draining/resolving old work and a new database. Credential rotation alone does not. Stop the worker before copying/backing up SQLite, preserve its WAL if present, encrypt backups and set retention for help desk content, transcripts and local records.
Recover without replaying unknown writes
uncertain-turn may include completed tool side effects: inspect the transcript before using restart JOB_ID --reviewed-side-effects. For uncertain-send, inspect native history and delivery. If already sent, use mark-sent JOB_ID; if confirmed never accepted, use retry-send JOB_ID --confirmed-not-sent to retry the cached answer. Run these through node --env-file=.env cli.mjs …; ownership is checked again before any send.
Monitor poll errors, API quota, failed delivery, disk and held-job age. Rotate/revoke credentials with the provider administrator and retest. For rollback, route/assign work to human support, resolve queued jobs, stop the relay and then revoke its credentials. Receiving remains in the help desk during relay outages; reconcile the native backlog against durable jobs before resuming.