Changelog

The changelog covers product changes to OpenTrain AI.

November 1, 2026

Changed

Withdrawn: the deprecation of time heartbeats sent with an API token other than a desktop device token

Updated October 6, 2026. The deprecation previously announced in this entry is withdrawn. Time heartbeats sent to POST /api/public/v1/time/heartbeat with an API token other than a desktop device token (such as a personal API token) keep working, and there is no sunset date. Responses do not carry Deprecation, Sunset, or Link headers for this. You don’t need to change your scripts or integrations.

September 4, 2026

Highlights this week

Every job now has an Agent Guide: one guidance document that every agent receives and keeps in sync automatically, with a human-published Playbook and a new employer workspace for publishing it and reviewing agent proposals. Instructions gain a safe suggesting mode, a restricted admin-balance workflow lets eligible platform administrators internally fund later milestones, employers get an Availability tab with a weekly heatmap, and the CLI and MCP get cleaner, more predictable output. Plus fixes for course Preview mastery states, editor stability, Work Email under load, and external job listings.

New

• Agent Guide for every job (CLI/MCP 0.27.0) — Every authorized agent managing a job through the API, CLI, MCP server, or SDK now receives the same Agent Guide: a read-only projection of the job’s typed context plus a human-published Playbook. Synchronization is automatic. A new session is directed to read the guide once, and after that agents fetch only the exact sections that changed between published versions. Agents propose Playbook changes; only humans publish, and nothing an agent proposes is followed until a human publishes it. Requires SDK 0.25.0+. See Agent Guide and Follow and improve the Agent Guide.
• Agent Guide workspace for employers — Job pages gain an Agent Guide tab where your team sees exactly what agents receive, writes and publishes the versioned Playbook from one shared draft, reviews agent proposals, and tracks version history and each agent session’s last sync. See Agent Guide workspace.
• Suggesting mode for job Instructions — Employers can now draft changes to an Instructions page as a suggestion instead of editing the live page. Review the changes side by side, then accept to update the live page or discard. If the live page changed in the meantime, the accept is refused and nothing is overwritten. See CLI: Job Instructions.
• Admin-balance milestones for platform administrators — Eligible OpenTrain platform administrators in the canonical admin organization can now create and internally fund a later milestone on a contract, or apply admin-balance funding to an existing eligible unfunded milestone in place, across the API, CLI, and MCP. The funding step moves $0 and creates no Stripe charge, ordinary employer funding is unchanged, and releasing payment still requires human authorization. See Create Admin-Balance Milestone and Fund Existing Milestone with Admin Balance.
• Availability heatmap for employers — Job pages now include an Availability tab with a weekly heatmap of the team’s shared schedules, so employers can see coverage at a glance. AI trainer schedule edits now save automatically. See Managing work.
• Job availability for employer agents — Agents can now run the full availability workflow from the API, CLI, and MCP: enable availability on a job, plan and confirm assignments, set the schedule time zone, and read the rolling 7-day schedule. Mutations are no-write previews until explicitly confirmed. See CLI Command Reference.

Updates

• Deterministic CLI output — Every non-stream CLI command now returns exactly one versioned result envelope on stdout, with --pretty as the explicit prose view. Usage failures exit 2, operation failures exit 1, and streams emit typed JSONL with one terminal event, so scripts can parse results reliably. See CLI Command Reference.
• Leaner MCP results — Job Operations, Instructions, and messaging tools now return a compact, deterministic summary instead of repeating the full payload, so agent context stays small. The complete validated result is unchanged in structured content, and credentials, links, and contact details are never mirrored into summaries. See MCP Tool Reference.
• More Instructions page icons — The page icon picker offers a larger catalog, and the Instructions page tree uses the full sidebar width and tracks your scroll position. See CLI: Job Instructions.
• Tidier Developer settings — The Manage platform app modal now uses a two-pane layout, and each API key lists its explicit restricted grants so it’s clear what a key can do. See Developer platform.

Fixes

• Course Preview mastery states — Employer course Preview now follows the mastery assignment policy and labels mastery states clearly. See Forms and quizzes.
• Editor stability — Pasting a YouTube link no longer crashes the Instructions and course editor, nested lists export to Markdown correctly, and editing no longer freezes on iOS and Android. See CLI: Job Instructions.
• Work Email under load — Inbox lookups stay fast on large accounts, and concurrent updates now retry safely instead of failing. See Use your Work Email.
• Folder action confirmations — Moving jobs into folders now shows a single confirmation, and only after the change is saved.
• External job listings — Listings from managed sources now always include a description, and non-English roles default to a basic English requirement so they match the right candidates. See Finding jobs.

September 3, 2026

Highlights this week

Job Messages gain semantic search so you can find any conversation by meaning, and AI trainers can now share a rolling 7-day availability schedule for each job. The New eval session dialog gets a cleaner design, Training course pages use the same editor as job Instructions, and fixes land for Instructions page reordering, Work Email timeouts, and job match delivery.

New

• Semantic search in Job Messages — Search a job’s messages by meaning, not just exact keywords. Press ⌘K (Ctrl+K on Windows) in the messaging panel to search across messages and people, then jump straight to the match in its conversation. See Messaging.
• Rolling 7-day work availability — AI trainers can now share their upcoming availability for each job from the job’s Manage tab. Availability covers the next 7 days in the schedule time zone saved for that job, so each job keeps its own schedule and employers can plan work around real hours. See Managing work.

Updates

• Redesigned New eval session dialog — Starting an eval session now uses a cleaner dialog: modes appear as selectable cards and run limits use tidier inputs, so setup is quicker to scan.
• Consistent editor in Training courses — Course pages now use the same editor as job Instructions, and the image upload control only appears where images are supported.

Fixes

• Instructions page reordering — Reordering pages in the Instructions sidebar works reliably again, including with a keyboard, and linked pages stay out of drag mode so they cannot be moved by mistake. See CLI: Job Instructions.
• Work Email timeouts — Work Email inbox operations no longer time out. See Use your Work Email.
• Job match delivery — Job match notifications now deliver reliably during busy periods.

September 2, 2026

Highlights this week

Every job now gets a Dynamic Work Catalog so multiple agents can define, claim, and share work safely, and Instructions pages gain automatic draft history with safe restore. Agents get authoring guardrails for Training courses and individualized retraining, plus new recruiting operations for reading job configuration, paging the talent census, and dispositioning proposals. Work Email addresses become readable, tracked time now finalizes when a contract ends, and fixes land across notifications, profiles, messaging, and the Instructions preview.

New

• Agent-readable visual proof for Instructions and quizzes (CLI/MCP 0.23.0) — Agents can now render the real Instructions workspace and Native Forms worker runner into desktop/mobile PNGs, ARIA snapshots, diagnostics, and provenance manifests. Direct evidence does not require a browser or Vercel account; short-lived OpenTrain-authenticated links are available for human review. Quiz previews create no attempts or learner state, and Instructions worker proof uses a designated real test contract. See Hosted Instructions visual proof and Hosted quiz visual proof.
• Dynamic Work Catalog for agents (CLI/MCP 0.23.0) — Every job now carries its own catalog of atomic work, defined by the team instead of a fixed platform list. Agents discover the catalog automatically during real work, record provisional work, propose reusable definitions from evidence, and claim work with leases so nothing is handled twice. Definitions activate only after a job manager approves them, and a per-job coordination policy controls how strictly claims are enforced. Available across the API, CLI, and MCP, with SDK 0.21.0+. See Dynamic Work Catalog.
• Draft history for Instructions pages — Every change to an Instructions page is now checkpointed automatically. Employers and agents can list a page’s history, compare any two versions, add manual checkpoints, and restore an earlier version safely: each restore records an undo checkpoint first and never overwrites concurrent edits. Available across the API, CLI, and MCP. See CLI: Job Instructions.
• Authoring guardrails for Training courses — Course validation now flags risky authoring patterns before publish: quiz-like content without an assessment, predictable answer positions, operational metadata in learner-facing copy, and overlong first checks. Publishing requires acknowledging outstanding warnings, and edits invalidate earlier acknowledgements. Contradictory mastery setups are rejected outright. Available across the API, CLI, and MCP. See Individualized retraining API.
• Typed authoring intent for retraining — Individualized retraining drafts now record an explicit authoring intent (individual mastery or standard employer review), and draft checkout, diff, push, publish, and Preview all preserve it. See Manage individualized retraining.
• Recruiting operations for agents (CLI 0.20.0) — Agents can now read owner job configuration, page through the talent census, and set proposal dispositions with a full audit history, across the API, CLI, and MCP. See Evaluate candidates.
• Readable Work Email addresses — New Work Email addresses are now based on your first name and last initial, so they are easy to recognize and share. Addresses that were already issued never change. See Use your Work Email.

Updates

• Tracked time finalizes at contract end — Hubstaff time now finalizes automatically when a contract ends, so final hours are recorded without manual follow-up. See Time tracking.
• Consistent job tabs for AI trainers — Job pages in the AI trainer dashboard now use the same underlined tab navigation as employer job pages, so the product reads the same on both sides. See Managing work.
• Course rollout visibility for agents — Active assignment rollouts now appear in agent-facing reads. See Course rollouts.
• Polished Instructions page icons — Subpages now default to a blank-document glyph, the icon picker matches OpenTrain branding, and page-title tooltips appear instantly without flicker. See CLI: Job Instructions.
• Cleaner proposal form and interview setup — The proposal form has a tidier layout with better-placed notices, and the mic and speaker check before the AI interview shows a mirrored self-view and clearer messages when a mic fails. See AI interview.

Fixes

• Instructions preview access — Customer preview links and Instructions previews no longer return errors after the September 1 incident.
• Notification redirects — Notifications now open the page they refer to instead of landing on the wrong surface.
• Profile editing — AI trainers now land on their editable profile from profile links, and agencies can correct their display name in settings.
• Verified tick in messaging — The verified tick now shows alongside the presence indicator instead of being hidden by it.
• Reviewer feedback triage — The triage view has clearer visuals and links directly to the work each piece of feedback is about. See Manage reviewer feedback and task checklists.
• Long dropdown menus — Long select menus scroll within the viewport again.
• Work Email reliability — Work Email now recovers safely when a primary address is missing, and waits for a usable name before issuing a first address. See Use your Work Email.
• Steady job tab counts — Count badges on employer job tabs now load without shifting the page layout.
• Performance under load — Database hot spots that could slow the platform during busy periods are fixed.

September 1, 2026

Highlights this week

Hosted visual previews let employers and agents see the real Training course player without a worker account, reviewer evaluations now capture balanced feedback instead of only concerns, Project To-dos can now unlock in order, and senders can edit or delete their own messages from the API. Plus quiz attempt evidence and time reconciliation for agents, page icons in Instructions, and fixes across employer jobs, proposals, candidate privacy, API key safety, and agreement signing.

New

• Hosted visual previews for Training courses — Employers and agents can now render the real worker course player remotely: capture desktop and mobile screenshots, run safe preview-only interactions, and share short-lived, observe-only review links. Previews never create learner progress, quiz attempts, or notifications. See Hosted LMS visual preview.
• Balanced reviewer evaluations — Reviewer feedback now starts with an overall assessment (Positive, Mixed, or Needs improvement) with optional anchored ratings, so praise gets recorded, not just problems. Feedback that needs no follow-up lands in a new Recorded lane without triage, and per-person summaries show rating distributions rather than averages. Available in the dashboard and across the API, CLI, and MCP. See Manage reviewer feedback and task checklists.
• Edit and delete sent messages from the API — Senders can now edit or delete their own messages through the API, CLI, and MCP. Every change keeps a revision history, and deleted messages read back safely instead of erroring. See Messages API.
• Quiz evidence, learner activity, and time reconciliation for agents (CLI 0.19.0) — Agents can now read immutable quiz attempt evidence, per-learner Training activity, and a job’s time-tracking reconciliation from the API, CLI, and MCP. See Developer platform.
• Ordered Project To-dos — A Project To-do can now be gated behind prerequisite to-dos or a time delay. Gated to-dos stay invisible to AI trainers until they unlock, and each unlocks exactly once. See Project To-dos.
• Exact mutation discovery for Job Operations — Agents can now query a canonical catalog of every available Job Operations mutation from the API, CLI, and MCP, so they know exactly what they can change before acting. See CLI: Job Operations.

Updates

• Sturdier hosted preview captures — Hosted preview captures that fail now report a typed reason with retry guidance and recover automatically, full-course preview audits carry verified assessment states, and the preview player can render quiz results safely without exposing answer keys. See Hosted LMS visual preview.
• Page icons in Instructions — Every Instructions page can now carry an emoji or icon, shown in a cleaner, more compact page tree. Icons are versioned with releases and readable across the API, CLI, and MCP. See CLI: Job Instructions.
• Stable deep links to Training content — Canonical job-scoped deep links now point to specific Training course content and always render the exact version they reference. See Forms and quizzes.
• Richer link previews — OpenTrain pages now show updated titles and preview images when shared in social and chat apps.

Fixes

• AI trainer API keys are read-only — API keys issued to AI trainers can no longer perform write operations. See Scopes and capabilities.
• Agreement signing — Embedded agreements now finalize promptly after signing, new signature requests are prioritized ahead of bulk processing so signers aren’t kept waiting, and e-signature Project To-dos recover automatically from temporary signing-provider failures. See Complete project to-dos.
• Employer jobs list — Jobs no longer drop out of the employer jobs list while it loads or filters.
• Hired proposals — Proposal statuses now stay aligned with contracts, so hired candidates show consistently across the Proposals tab and proposal messages. See Review proposals.
• Candidate privacy on agent surfaces — Candidate contact details are now redacted from agent-facing API, CLI, and MCP responses. See Evaluate candidates.
• Candidate data reliability — Candidate reads stay reliable under heavy load, and malformed legacy profile fields no longer cause errors. See Developer platform.
• Simpler feedback for AI trainers — The AI trainer feedback view is simpler, and quality feedback reports can now be shared without naming the reviewer.
• Employer name casing — Employer names now keep their exact capitalization everywhere they appear.
• Contract status in messages — The message panel header now labels the status badge “Contract: Active”, so it’s clear the status refers to the contract.

August 31, 2026

Highlights this week

Shared Job Context now sets itself up automatically the first time an agent touches a job, and agents stay coordinated through a new real-time sync layer. Quiz coaching, task checklists, and reviewer feedback reach full API, CLI, and MCP parity, interview scoring gets a more trustworthy rubric, and the Instructions editor and sign-in pages are redesigned. Plus fixes across payouts, Work Email, interviews, and login.

New

• Automatic Shared Job Context setup — The first job operation on a job now starts a guided setup: OpenTrain detects what’s missing, proposes evidence-backed defaults, plans every change with zero writes, and applies only after explicit approval. One agent owns setup at a time, and teams can defer setup for everyone. Available across the API, CLI, and MCP. See Set up shared job operations.
• Real-time agent sync — Every job-scoped operation now returns a sync envelope with acknowledged cursors and durable agent sessions, and stale writes are caught before they land, so multiple agents can work the same job without losing updates. See Stay in sync.
• Quiz coaching for agents (CLI 0.14.0) — Employers can now run the full quiz coaching loop from the API, CLI, and MCP: review submissions, comment, request redos, and accept or send back work, with feedback images and exports. AI trainer agents can respond in kind. See CLI: Project To-dos.
• Checklists and reviewer feedback for agents — Task checklists and reviewer feedback can now be authored and managed across the API, CLI, and MCP, matching what the dashboard offers. See Manage reviewer feedback and task checklists.
• Anti-guess quiz policies — Employers can add escalating cooldowns, attempt limits, and restart scopes to quizzes, so mastery can’t be brute-forced. Enforced in the dashboard, preview, AI trainer player, and across the API, CLI, and MCP. See Forms and quizzes.
• Interview-only score in the API — Proposals, transcripts, and candidate dossiers now expose the interview-only score alongside the composite score. See Developer platform.
• Candidate resume access via the API — Fetch candidate resumes through the API, CLI, and MCP with short-lived signed download links, gated by a dedicated scope and fully audited. See Evaluate candidates.
• Work Email lookup for agents — Look up an AI trainer’s managed Work Email address from the API, CLI, and MCP, scope-gated and audited. See Privacy and Work Email.

Updates

• More trustworthy interview scoring — Scores are now anchored to verified correctness, with deterministic caps for confirmed errors, credit for equivalent tools, and automatic downgrades of contradictory recommendations. All downstream thresholds are recalibrated to match. See Review proposals.
• Redesigned Instructions editor — A rebuilt editing experience with page-tree navigation, a modern search palette with more relevant results, and redesigned video insertion backed by a reusable video library with upload recovery. See CLI: Job Instructions.
• Course lessons share the Instructions editor — Training course lessons now use the same rich-document editor as Instructions, with fuller formatting controls, faithful math rendering, and identical preview and player output. See Forms and quizzes.
• Choose who receives a Project To-do update — Publishing an update to a Project To-do now includes an audience picker with an impact preview and notifications, in the dashboard and the API. See Project To-dos.
• Exact-page Instruction links for agents — Agents can read canonical, stable deep links to specific Instructions pages across the API, CLI, and MCP. See Manage job instructions.
• Redesigned sign-in — Refreshed sign-in and sign-up pages with better mobile layouts and clearer messaging for AI trainers, employers, and agencies.

Fixes

• Payouts — Expired Stripe payout setup links now renew automatically instead of dead-ending, and paid earnings display correctly on ongoing jobs. See Payouts.
• Work Email — Issued addresses are now permanently immutable, new addresses reveal less of your name, and API responses carrying Work Email data are never cached. See Use your Work Email.
• Voice interview audio — Interview audio now buffers and recovers after network hiccups, so answers no longer stutter or cut out on slow connections. See AI interview.
• Interview-only score display — The interview-only score now shows consistently in the People table, messaging, and mobile proposal views. See Managing candidates.
• Verification reminders — Identity verification reminders stop once your verification is in review or complete. See Review proposals.
• Login reliability — Signing in stays reliable under heavy load.
• Pinned jobs — Pinned jobs now always stay above folders in the employer jobs list.
• Project To-do assignments — Concurrent assignment changes now converge to one authoritative assignment, so no duplicate assignments or repeated invites.
• Developer platform — Shared Job Context reads now return a typed, retryable error on timeout instead of failing opaquely, bulk Instructions applies are atomic with true dry runs, and training progress tracking in Shared Job Context is corrected. See Shared Job Context.

August 30, 2026

Highlights this week

Shared Job Context gains attention signals with claim-before-work, typed job policies, and a generated Job Operating Guide, all available across the API, CLI, and MCP. Plus sturdier individualized retraining automation for agents.

New

• Attention signals in Shared Job Context — Each job’s shared context now carries a ranked queue of deterministic attention signals for conditions that may need investigation. Agents and employer team members triage signals with acknowledge, snooze, or dismiss, and must claim a signal before turning it into owned work, so nothing gets handled twice. Available across the API, CLI, and MCP. See Shared Job Context.
• Typed job policies — Record typed policies on a job, read the effective policy set, and verify external enforcement with evidence. Policy history preserves attribution and lineage, so agents always act on the current rules. See CLI: Job Operations.
• Job Operating Guide — OpenTrain now generates a read-only operating guide for each job from its current objective, workflow, effective policies, approved exceptions, and fresh attention signals. It is the shared AGENTS.md-equivalent for that job, readable from the API, CLI, and MCP. See MCP: Job Operations.

Updates

• Sturdier retraining automation — Individualized retraining is easier to drive with agents: review reads return complete, canonical grading context, CLI errors are machine-readable under --json, and long status and roster lists page deterministically with resume cursors. See Manage individualized retraining.

August 29, 2026

Highlights this week

Training courses on every job with grading feedback and individualized retraining, Shared Job Context for agents, reviewer feedback and task checklists, in-product agreement signing, native video in Instructions, and live presence in messages. Plus reliability fixes across payouts, Work Email, verification, messaging, and interviews.

New

• Training courses on every account — Employers can now build Training courses on any job: a per-job Training tab with course authoring, private video, and quiz assessments, assigned to AI trainers as Project To-dos. AI trainers complete courses in a dedicated Training area. See Complete project to-dos.
• Grading feedback and resubmission — Employers can grade training submissions question by question, add written feedback, and return work for revision. AI trainers see reviewed results and can revise and resubmit. See Forms and quizzes.
• Individualized retraining — Employers can assign a single AI trainer a targeted retraining course from a documented quality gap, preview it before assigning, and track it as a Project To-do, with full API, CLI, and MCP support. See Manage individualized retraining.
• Shared Job Context for agents — One evidence-backed operational view of a live job, so coding agents and employer team members can triage, act, and hand off safely across the API and MCP. See Shared Job Context.
• Reviewer feedback and task checklists — Jobs can now opt in to per-task checklists for AI trainers and a reviewer feedback workflow that delivers reviewer notes directly to the AI trainer who did the work.
• Sign agreements in-product — Agreement Project To-dos now open in an embedded signing view, so AI trainers sign without leaving OpenTrain. An email fallback covers cases where embedded signing isn’t available. See Complete project to-dos.
• Native video in Instructions — Upload videos directly into job Instructions with private, in-page playback, including uploads from the CLI. See CLI: Job Instructions.
• Live presence in messages — Job channels now show who is currently online. See Job channels.

Updates

• People table: to-do progress and faster paging — Project To-do progress now rolls up into each job’s People table, with an expandable drill-down per person. Large rosters load with server-side paging and support batch edits. See Managing candidates.
• Redesigned Project To-dos control center — The employer Project To-dos panel has a cleaner layout with unified review and progress views. See Project To-dos.
• Instructions editor polish — A redesigned page tree, release bar, and publish modal, share links that open the exact instruction page, and images that open in an in-app lightbox. See Developer platform.
• One verification verdict everywhere — A candidate’s identity verification status now matches across proposals, profiles, dossiers, and the People table, and AI trainers see verification events in their messages. See Review proposals.
• Proposals from job messages — Employers can jump from a job message directly to the sender’s proposal. See Review proposals.
• Employer account menu — A refreshed account menu opens from the employer sidebar avatar.
• Voice interview setup redesign — The mic and speaker setup before the AI interview has a clearer design. See AI interview.

Fixes

• Payments — Milestone approval no longer waits on payout delivery, payouts recover automatically after a safety reversal, Stripe payout state is verified before settlement, and contracts with pending payouts can no longer be closed unsafely. See Payouts.
• Work Email delivery — Mailbox assignments can no longer change unexpectedly, so Work Email delivers reliably. See Use your Work Email.
• Realtime messaging — Job and inbox messages update in real time again, the composer is more reliable, and message counts no longer misreport while loading. See Job channels.
• Interview recordings — Long answers are preserved in full, truncated clips recover safely, and employer audio playback is stable. See AI interview.
• Hubstaff and provider offboarding — Hubstaff access is revoked when a to-do is removed or a contract ends, setup verification is more reliable, and one-time setup assignments are preserved. See Set up Hubstaff.
• Training reliability — Private asset uploads, quiz focus behavior, module-by-module learner progression, and AI trainer video access after course updates all work correctly again.
• Faster page loads — Cold-start slowdowns and unbounded Job Messages loading are fixed.
• Employer workflows — Profile save errors now surface clearly, job posting details match the review step, contract search covers all job contracts, and hire submission no longer gets stuck.
• Onboarding accuracy — Resume-based country of residence inference and phone country correction work correctly during signup. See Profile setup.
• Developer platform — Talent search and token authentication are more reliable under load, and MCP now reports its installed runtime version. See Developer platform.

August 25, 2026

Highlights this week

Forms and quizzes for project onboarding, a People table for employer jobs, Work Email for AI trainers, Hubstaff time tracking, ID verification on proposals, and agent-ready job Instructions across the API, CLI, and MCP. Plus CLI 0.5.2, in-job quiz results, agreements that prefill engagement terms, and reliability fixes across verification, job drafts, and Instructions.

New

• Forms and quizzes — Employers can now build forms and quizzes in a reusable library or directly inside a job, and attach quizzes to Project To-dos so AI trainers complete them before starting work. See Complete project to-dos.
• Signed agreements before work — Employers can require e-signature agreements as Project To-dos, so AI trainers sign in-product before the project’s work area opens. See Complete project to-dos.
• People table for employer jobs — A spreadsheet-style People view on each job with saved views, filtering, and typed custom fields for tracking candidates and hires in one place. See Managing candidates.
• Work Email for AI trainers — Verified AI trainers now get a managed @opentrain.work inbox for work-related email, with an unread badge on the dashboard so new mail is easy to spot. See Privacy and Work Email.
• Hubstaff time tracking setup — A verified Hubstaff setup flow tied to your OpenTrain Work Email account, with setup progress visible to both AI trainers and employers. The Hubstaff Project To-do now walks AI trainers through setup step by step. See Set up Hubstaff.
• ID verification on proposals — Employers see identity verification status directly in proposal lists and details, applicants see up front when a job requires ID verification, and automatic reminders nudge candidates to finish. See Review proposals.
• Agent-ready job Instructions — Job Instructions are now canonical across the API, CLI, and MCP, with first-class workflows for coding agents. See Developer platform.
• Instructions CLI 0.5.2 — The opentrain CLI ships the complete Instructions command group, adding guarded image uploads and command-specific help. See CLI: Job Instructions.
• Quiz results inside the job — Employers now review and release quiz results directly from the job’s Project To-dos panel, so results stay next to the work they gate. See Complete project to-dos.

Updates

• “AI Trainer” naming everywhere — The product now consistently says “AI Trainer” in place of “Freelancer” and “Labeler” across proposals, contracts, and settings.
• Sent and Hired proposal tabs — The employer Proposals tab adds Sent and Hired tabs so it’s easier to track candidates through the funnel. See Review proposals.
• Redesigned Developer settings — API keys now appear above Platform apps in Employer Settings → Developer, with a compact create-key modal and app details collapsed by default. See Developer platform.
• Better messaging — Image attachments open in an in-app lightbox with previous/next navigation instead of a new browser tab, and URLs in messages are now clickable links. See Job channels.
• Owner-only channel deletion — Group channel owners can now delete channels, and deleted channels disappear immediately for all members. See Job channels.
• Shared job links open the job page — Tracked job links now land visitors on the job page first, and the Create Account to Apply button carries them through signup and back to the job.
• Agreements prefill engagement terms — E-signature agreement to-dos now generate documents with the engagement’s specific terms filled in automatically. See Complete project to-dos.
• Easier Project To-do management — Employers can now search the assignee picker when assigning a to-do, and reviewing submissions from the Project To-dos panel is smoother.
• Compact Add capture source form — The capture area’s Add source form now groups fields into sections and fits smaller screens, so the footer actions stay visible.
• Reorder job tabs — Employers can now arrange the tabs on each job’s page (Proposals, People, Messages, and more) in the order they prefer. The order is saved per person for each job. See Managing candidates.
• Rebuilt New to-do modal — The employer modal for creating Project To-dos has a cleaner, easier-to-scan layout. See Complete project to-dos.

Fixes

• Work Email privacy — Work Email no longer exposes AI trainers’ personal email addresses to employers.
• Fairer credential matching — Matching now recognizes equivalent international credentials, so qualified candidates are no longer rejected from jobs requiring a specific credential they hold an equivalent of. See Finding jobs.
• Payouts during setup — Payouts requested before payout setup finishes now queue and release automatically once setup completes, instead of failing. See Payouts.
• Employer navigation — The Contracts tab and My Jobs navigation work reliably again, and job lists no longer show a false empty state while loading.
• Agency proposals — Agencies can submit proposals from the job detail page again.
• Ended-contract messages — Direct messages from ended contracts no longer appear active in a job’s Messages tab, and the AI trainer inbox now defaults to showing all conversations.
• Employer job drafts — Creating and editing job drafts works reliably again in the job post flow.
• ID verification accuracy — Verification results now keep the strongest valid evidence on record, so verified candidates no longer show as unverified. See Review proposals.
• Instructions reliability — Instructions pages no longer appear blank while a job first syncs, agent authoring and collaboration sessions are more robust, and clean CLI installs work out of the box.

June 24, 2026

Highlights this week

A practice question before the scored AI interview, in-call volume controls for candidates, smoother Stripe payout resumes, and reliability fixes across voice interviews, Find Jobs, signup, and support.

New

• Warmup practice question — AI trainers now get a short, unscored warmup question before the scored AI interview so they can settle in, check their audio, and hear what a real prompt sounds like before it counts. See AI interview.
• In-call volume control and “can’t hear” helper — The AI interview panel now includes an in-call volume slider and a “Can’t hear?” helper that walks candidates through quick audio fixes mid-interview. See AI interview.
• Resume in Stripe from payout settings — When a payout account is awaiting Stripe verification, AI trainers now see a Resume in Stripe button in payout settings to jump straight back into Stripe and finish onboarding, alongside the existing Refresh Status option. See Payouts.
• Payment-policy support auto-reply — Support questions about payment-related privacy and data handling now get an immediate, on-topic auto-reply with the right policy info instead of waiting in the queue.

Updates

• Find Jobs only shows currently accepted external roles — Find Jobs now filters external listings to those backed by the latest accepted job audit, so AI trainers don’t see external roles that have since been revised or pulled. See Finding jobs.
• Sharper signup terms hint — On the signup role-selection step, a clear inline hint now tells you to agree to the Terms of Service and Privacy Policy before continuing, instead of a silently disabled button. The selected role is also remembered if you navigate back.

Fixes

• Voice interview answer detection — Endpointing no longer cuts AI trainers off mid-sentence, typical answer turns end faster, and a new inactivity backstop keeps interviews moving when transcript signals stall. See AI interview.
• Voice quality watchdog — A background voice-quality watchdog now monitors interview audio for truncated answers and other quality issues so problems get caught and addressed faster, with a recalibrated threshold to avoid false alarms.
• Voice interview preflight recovery — The mic and speaker preflight check now recovers cleanly from transient browser worker errors instead of getting stuck before the interview starts. See AI interview.

June 22, 2026

Highlights this week

A refreshed employer Proposals tab for reviewing and comparing candidates.

Updates

• Refreshed employer Proposals tab — A cleaner layout for reviewing and comparing candidate proposals, including an updated bulk message composer. See Review proposals.

June 19, 2026

Highlights this week

Smoother AI voice interviews, more reliable onboarding for AI trainers, and reliability fixes across employer settings, payouts, and messaging.

New

• Stripe preview-country payout notice — AI trainers in preview-supported payout countries now see a clear in-product notice during payout setup explaining what’s available today and what’s coming. See Payout country availability.
• Employer payment settings access guard — Access to employer payment settings is now gated by stricter role checks, so only authorized teammates can view or change billing details. See Team members.

Updates

• Steadier AI voice interviews — Voice playback uses a deeper jitter buffer with underrun concealment and a speaker pre-check, and a new compression path reduces bandwidth for cleaner audio on slower connections. Interviews also recover better from brief reconnects without ending early. See AI interview.
• Resume autofill for data-labeling roles — Resume parsing now extracts data-labeling experience and tooling more accurately, so AI trainer profiles autofill with the right skills. See Profile setup.
• Stronger matching for specialized roles — The matcher now enforces primary evidence requirements for specialized job plans, so AI trainers see specialized roles that genuinely match their proven skills. See Finding jobs.

Fixes

• AI trainer profile completion — Country corrections, Brazil phone validation, state selection, skill dropdown completion, and profile overview persistence all work reliably during onboarding. See Profile setup.
• Benin country selector — Benin is now available in the account location country selector.
• AI interview reliability — Blank voice turns and stale reconnects no longer end interviews prematurely, and voice recordings upload reliably at the end of a session. See AI interview.
• Phone confirmation rate limits — Phone confirmation surfaces a clear message when SMS rate limits are hit, instead of a generic error.
• Support launcher overlap — The support help launcher no longer overlaps the message composer on smaller screens.

June 8, 2026

Highlights this week

Employer project sharing, better job discovery for language and specialist roles, and reliability fixes across AI trainer onboarding, contracts, billing records, and messaging.

New

• Share projects with teammates - Employers can now open Share from a job or folder header, add existing teammates who should have project access, and review the access list in Job Settings. Owners keep access, and non-owner teammates can only share projects they can already open. See Share projects with teammates.

Updates

• More accurate job matching for language and specialized roles — The matcher better distinguishes language-driven data tasks, professional and advisory roles, and domain-anchored requirements, so AI trainers see jobs that fit their actual skills. See Finding jobs.

Fixes

• AI trainer resume autofill — Resume parsing and profile autofill recover gracefully from timeouts and parser issues during onboarding. See Profile setup.
• Ended contract compensation view — Compensation details on ended contracts load reliably for authorized viewers. See Managing work.
• First milestone refund invoicing — Employer billing records now stay accurate after a first-milestone refund. See Invoicing and payouts.
• Messaging spam protection — OpenTrain blocks known scam messages before they reach job-channel inboxes. See Job channels.

May 30, 2026

Highlights this week

Pay-per-label contract switches, clearer job discovery, and reliability improvements across core hiring and work flows.

New

• Switch active contracts from hourly to pay-per-label — Employers can now change a live hourly contract to pay-per-label for future milestones. Existing funded milestones keep their original hourly terms. See Milestones.

Updates

• Find Jobs improvements — Job cards are easier to scan on desktop and mobile, with clearer descriptions, badges, and application status. See Finding jobs.
• Contract and messaging filters — Ended contracts are easier to spot and filter across contract lists, inboxes, and job messages. See Managing work.

Fixes

• Message edit expiry — Messages edited after the edit window now show a clear expiry notice instead of a generic error.
• Job search, profiles, and onboarding reliability — Job search, job details, employer dashboards, profiles, AI trainer onboarding, and resume autofill are more stable.
• Public job detail header — Public job detail pages now have cleaner header spacing and layout.

May 28, 2026

Highlights this week

Safer proposal messages, agency marketplace trials, provider invite access, and public job detail improvements.

New

• Proposal message safety screening — Proposal and bulk messages are screened for abusive content before delivery, helping keep conversations safe. See Proposal Messages.
• Agency marketplace free trial — Agencies can self-serve a 30-day free trial to access the marketplace without entering payment details up front.
• Provider invite access — Employers can invite external annotation providers directly into a job, granting scoped access to candidates and deliverables without sharing their full workspace.

Fixes

• AI trainer dashboard stability — The AI trainer dashboard is more reliable when contract data is temporarily unavailable.
• Public job detail — Candidates now see clearer job status before applying.