Changelog
What's new.
Product updates, infrastructure improvements, and new integrations — as they ship.
Releases
- Product
Public Help Center, campaign audience segments, and trademark
- Public Help Center: every workspace gets a branded Help Center at /help/[workspace-slug]. Articles are published directly from the knowledge base, rendered with a minimal markdown subset (headings, bold, inline code, links, lists — fenced code blocks are not supported), and accessible to visitors without login.
- Campaign audience segments: three new targeting segments for proactive campaigns — Dormant Visitors (no conversation in 30 or 60 days), Low-CSAT Visitors (CSAT score ≤2 stars in the last 90 days), and New Visitors (first conversation within the last 7 days). Segments are evaluated at send time against the live conversation history.
- Trademark: VELOR PLATFORMS® federally registered on the U.S. Principal Register (Reg. No. 8,321,848, Class 042, June 30, 2026).
- Security hardening (August): Help Center article delete now detects silent zero-row deletes; audience enum coverage extended to all campaign edge cases; unbounded selects replaced with explicit limits across Help Center endpoints; org_id guard added to all Help Center write paths.
- Product
Connector direct dispatch, widget branding, and industry pages
- Connector direct dispatch: Zendesk, Shopify, and Stripe connectors now call their APIs directly from the Worker — customers paste API keys into settings and the agent has live customer context immediately. No adapter infrastructure required.
- Widget custom branding: accent color, business name, and welcome message are now reflected on the embedded chat widget. Trial workspaces see their configuration applied in real time without a redeploy.
- Proactive messaging restored: a schema-drift bug was silently disabling proactive messaging across all workspaces. Fixed and re-enabled — scheduled and event-triggered outbound messages now fire correctly.
- Industry pages: dedicated pages for Fintech & Lending, HR/Payroll, and Staffing & Recruiting showing exactly how the guardrails, actions, and KB retrieval are configured per vertical. Linked from the footer.
- Health probe reliability: OpenAI reachability probe timeout raised to 4 seconds with one retry — transient edge latency no longer falsely reports OpenAI as down and tips the health endpoint to degraded.
- Infrastructure
Circuit breakers, trial lifecycle, and KB retrieval tuning
- Circuit breakers on every Anthropic call site: all inference paths now record success/failure and bail early when the breaker is open. Prevents cascading timeouts under LLM API degradation — the agent returns a graceful fallback rather than hanging.
- KB retrieval threshold lowered from 0.7 to 0.45: the semantic similarity cutoff was too conservative for short-embedding models, causing grounding to silently fall back to the LLM's training data. Lowering it restored retrieval for paraphrased queries. Superseded on 2026-08-25: 0.45 sat inside the real signal band (correct answers land at 0.35–0.60 cosine) and was itself dropping correct grounding, so the live cutoff is 0.30. Do not set 0.45 from this entry.
- Trial lifecycle improvements: day-5 rescue drip email, low-CSAT alert at day 3, first-conversation milestone email, and a dashboard KB draft nudge for workspaces that haven't published a knowledge source. Resolution-rate metric added to the admin panel.
- Onboarding polish: widget greeting no longer shows a blank bubble if the system prompt hasn't been configured yet; falls back to a sensible default. Trial banner now shows exact days remaining rather than a generic 'trial' label.
- Async local storage lazy-init fix in the Cloudflare Worker bundle: eliminated a startup error that caused the first request after a cold start to fail silently on Wrangler environments.
- Product
Voice IVR studio, white-label custom domains, and Slack inbound
- IVR studio: build interactive voice response menus from the dashboard. Drag-and-drop step builder supports collect-digit, speak, route-to-AI, route-to-queue, transfer, and voicemail steps. Menus are versioned; publishing a new draft is atomic.
- Voice persona controls: set tone (professional, friendly, casual), style notes and the ElevenLabs model per phone number. Speech rate, pitch and pause timing are NOT configurable — this entry claimed them and four preset profiles, and neither shipped; the synthesis call sends stability, similarity boost and style only.
- Custom domains: any workspace can verify a hostname it owns (e.g. support.acme.com) with a TXT record (velor-verify=…), and a verified hostname is added to the widget's allowed origins. Velor does not serve the widget script or the dashboard from a customer hostname — there is no routing or certificate for one — and an earlier version of this entry said a CNAME would route it. This entry previously named the wrong verification record, promised that we provision the certificate, and offered email links on your domain behind a Growth/Enterprise gate. We do not issue the certificate, email links are still built from the platform URL, and there is no plan gate.
- Slack native inbound: connect a Slack workspace so that messages in designated channels are processed by the AI agent. Replies thread back into Slack. Escalations route to the inbox with full Slack thread context attached.
- Copilot panel: human agents handling escalated conversations see real-time AI suggestions — draft replies, retrieved KB passages, and available action calls — in a side panel. One-click to accept a suggestion; editable before sending.
- Product
Skills system, quality rubrics, and BYOK
- Skills system: package reusable agent behaviors (a greeting script, a returns policy handler, an upsell sequence) as versioned skills. Assign one or more skills to a named agent. Multiple skills are concatenated into the prompt in list order. There is no conflict detection and no last-writer-wins rule — two contradictory skills both reach the model, so keep overlapping instructions out of skills you assign together.
- Quality rubrics: define custom scoring dimensions (accuracy, tone, policy compliance, escalation appropriateness) with per-dimension weights and thresholds. Rubrics are per-workspace and applied to every scored conversation. Reviewer verdict (Agreed / Partial / Disagreed) is now captured when closing a quality review.
- Customer-managed encryption keys (BYOK): AWS KMS, GCP KMS and HashiCorp Vault key adapters are implemented but are not yet applied to stored data. This is on the roadmap and is not available today. It would cover encryption at rest; it is not a bring-your-own LLM key — model inference runs on Velor's own provider accounts, and Anthropic is disclosed as Velor's sub-processor for every customer in the DPA.
- MCP server registry: register a Model Context Protocol server over HTTP and sync its tool catalog — names, descriptions and parameter schemas are fetched from the server itself, behind an SSRF guard, when an admin runs a sync. Once a server is synced and enabled, the agent can call its tools alongside custom actions, through the same controls: the Actions kill switch, a signed-in visitor, a confirmation before every call, argument and result sanitising, rate limits and an audit record. A server whose last sync failed contributes no tools. Teammate approval of an MCP call is not available yet.
- A/B experiments: run split tests on system prompt addendum, tone, or escalation threshold. Traffic is split by CONVERSATION id, so a returning visitor can see a different arm in a later conversation. Results surface at /dashboard/experiments with statistical significance indicators. KB retrieval config is not an experiment type, and this entry claimed it was.
- Infrastructure
Platform-wide security audit and reliability hardening
- 100-fix security audit pass: every API route, server action, and cron handler reviewed. Fixes include: SSRF guards on all user-supplied URLs (DNS-resolving check, not just prefix match), org-scope enforcement on every write, XML barriers in all LLM prompts to prevent prompt injection, body-size caps on all public endpoints, structured error responses that never leak raw DB messages.
- Rate limiting extended: chat poll, search, evals, assist/summarize, proactive analytics, and SCIM endpoints now rate-limited per user or IP. If the limiter's own store fails, the request is denied and a warning is logged rather than passed through; no alert is raised. This entry previously said limiter failures alert.
- Webhook secrets encrypted at rest using AES-256-GCM. Existing plaintext secrets migrated on next rotation.
- UUID validation added to every route and server action that accepts a resource ID. Malformed IDs rejected before hitting the database.
- Monitor digest alerts: after each monitor evaluation cycle, org owners receive a digest email of flagged conversations with reason snippets and direct links. 4-hour per-org cooldown prevents alert fatigue.
- Trial expiry reminders: org owners receive personal emails at 7 days and 1 day before trial end — no automated checkout, just a reply-to-continue flow.
- Security
Audit log, API keys, and SAML/SSO
- Audit log: append-only record of every admin action — API key creation and revocation, member changes, SSO configuration, settings updates, knowledge source changes. Visible to owner and admin roles. Retention defaults to 30 days and is configurable up to 365; this entry claimed 12 months, which is the ceiling rather than the default.
- Public REST API (v1): programmatic access to conversations and analytics via scoped API keys. Bearer token auth. Keys are scoped: conversations:read, analytics:read, knowledge:read, export:read. Full key shown once at creation; SHA-256 hash stored.
- API key management UI in Settings — create, name, scope, and revoke keys. Confirmation step before revocation. Last-used timestamp updated on every authenticated request.
- SAML 2.0 / SSO configuration: schema and setup wizard for Okta, Microsoft Entra ID (Azure AD), Google Workspace, OneLogin, and any SAML 2.0-compatible IdP. Domain enforcement sends addresses on the configured email domain to the IdP. DNS ownership verification of the domain was not implemented at this release; it has since been added (a TXT record is required before enforcement can be turned on).
- Product
Proactive campaigns and multilingual support
- Proactive campaigns: compose and send outbound messages to visitor segments from the dashboard. Trigger on events, schedule by time, or send immediately. Delivery, open and click rates tracked per campaign. Reply and conversion tracking do not exist — this entry claimed them.
- Multilingual: the agent detects visitor language on the first turn and responds in kind for the rest of the conversation. On for every workspace with nothing to configure; English fallback when detection is inconclusive.
- Language addendum in system prompt ensures the model never reverts to English mid-conversation, even when KB content is English-only.
- Campaign analytics surface in the Insights dashboard alongside resolution and escalation metrics.
- Product
Quality scoring and conversation explorer
- AI-powered quality scoring: every resolved conversation is scored 1–5 on helpfulness, accuracy, tone and resolution. Add escalation judgment, policy compliance or your own dimensions with a configurable rubric — the standard rubric includes all six.
- Score thresholds trigger automatic review queue routing — conversations below a set threshold surface in a dedicated review queue for your team.
- Conversation explorer: full-text and semantic search across all conversations. Filter by score, channel, escalation reason, agent, date range, or visitor.
- Bulk-review mode lets QA leads score a batch of conversations, leave notes, and track calibration over time.
- Product
Visitor memory and named agents
- Visitor memory: persistent facts about each visitor (plan tier, previous issue patterns, stated preferences) are surfaced to the agent at the start of each widget chat and voice call. Facts are stored per visitor profile, reviewable and editable in the dashboard.
- Named agents: create multiple agents with distinct names, personas, and system addenda. Route by channel, widget embed, or incoming trigger. Each named agent has independent KB access, action scope, and guardrail config.
- Agent configuration expanded: tone presets (professional, friendly, concise, formal), escalation trigger lists, forbidden topic lists — all configurable per named agent without touching the base config.
- Memory is read on the widget, voice, SMS, WhatsApp, email and Slack, and every channel receives the same formatted context. Voice, SMS and WhatsApp share one memory per phone number; email memory is keyed by the sender address; the widget by its visitor; Slack by the Slack user, and only for a thread whose Slack profile email could be read, so that a request to erase a person by email reaches it. Memory is not linked between a person's email, phone, Slack and widget identities.
- Product
Evaluations framework and playground
- Eval framework: run any conversation set against the agent offline. Score each turn — did the agent retrieve correctly, invoke the right action, escalate when it should have resolved? Results include per-turn pass/fail with reason codes.
- Playground: test agent responses live in the dashboard against any named agent or config variant. No production traffic affected. Useful for testing KB changes before publishing.
- Eval history tracks run diffs so you can see exactly what changed in retrieval quality or escalation rate between two config versions.
- Eval API endpoint for CI integration: run evals on every KB or prompt change as part of your deployment pipeline.
- Product
CSAT, live assist, and inbox handoff
- CSAT collection: post-conversation satisfaction survey sent via email (HMAC-signed one-click score links) and surfaced in the chat widget as an inline prompt. Score distribution and trend in the Insights dashboard.
- Inbox: escalated conversations route to a unified inbox. Assign to teammates, add internal notes visible only to your team, and hand back to the agent when the human is done.
- Live assist: when handling an escalated conversation, human agents see the full transcript and retrieved KB chunks in-context. This entry also claimed one-click action execution for humans; no inbox button runs an action, and destructive actions the agent proposes wait for a teammate's approval in the dashboard.
- CSAT email links are time-bounded (30-day expiry) and signature-verified server-side — no anonymous score stuffing. This entry previously said 72 hours.
- Infrastructure
Cloudflare Workers deployment and security hardening
- Full deployment to Cloudflare Workers via opennextjs-cloudflare.
- Row-level security audit: every customer table verified with Postgres policy-level RLS, so queries made with a customer's session cannot read another tenant's rows. Server-side paths that use the service-role key bypass RLS by design; on those paths the organization filter in the query is the boundary. This entry previously called cross-tenant access impossible at the database layer.
- HMAC-SHA256 signed outbound calls: action calls carry Velor-Signature-256 (v1=<hex>) and Velor-Timestamp. Registered actions are always signed; actions made in the no-code HTTP builder are signed when they have a signing secret — every builder action created since 2026-10-09, and older ones once an admin rotates the secret. Webhook deliveries are a separate system and carry x-velor-signature. Both cover the request timestamp + body, and the timestamp lets your server reject replays outside five minutes. SSRF protection added at dispatch. This entry previously said every action call carries an X-Velor-Signature header.
- Product
Chat widget, email channel, and the core agent runtime
- Chat widget: embeddable with a single script tag. Supports visitor identification (HMAC-signed userId + traits), custom theming, and mobile-responsive layout. Streams agent responses token-by-token.
- Email channel: inbound emails forwarded to a Velor intake address are processed by the agent and replied to in-thread, with In-Reply-To / References headers maintained for Gmail, Outlook, and Apple Mail threading. The intake address is set up with you rather than switched on yourself.
- Knowledge Base: articles, FAQs, and internal docs chunked, embedded, and indexed on upload. Semantic retrieval on every turn. Cited passages visible in the transcript.
- Custom actions: define API tool calls via JSON schema. HMAC-signed dispatch (builder actions created before 2026-10-09 only once their secret is rotated), schema-validated inputs and configurable timeouts. Call rate limits are fixed, not configurable (this entry said they were). Destructive actions need the visitor's confirmation on a card in the chat widget, or a teammate's approval on every other channel.
- Guardrails: scope enforcement, PII handling rules, escalation triggers (confidence, sentiment, content class), and destructive-action confirmation — all configurable per workspace.
- Escalation: clear criteria route conversations to the inbox with full transcript, customer history, and retrieved KB context attached. Reason codes surface in the dashboard.
Have feedback on what we should ship next? Talk to us →