Changelog

Platform

Runs are told what is missing while they can still fix it

When a run is about to finish without the step that proves its work, OpenHelm now tells it so before it ends, rather than failing it afterwards.

• A run that wrote files into your repository but never committed them is shown exactly which files are at risk.
• It is told plainly to make the commit, not describe one.
• Monitoring jobs can require proof they actually looked before reporting "nothing to do".

Grow

Blog posts are checked before they can break your site

The blog publisher now reads each post back before committing it, and publishing is split into smaller chained jobs so every step can be checked on its own.

• Frontmatter is re-read after writing, because one unquoted colon can fail the whole site build.
• New posts copy the formatting your existing posts already use rather than inventing their own.
• A separate job confirms the post is live over HTTP and logs it.
• Backfilling posts published before OpenHelm arrived is its own weekly job.

Understand

Category screens chart what your agents record

A template can now declare the charts and tables its category screens should show, so the numbers your agents collect appear without any setup.

• Growth shows search performance, content performance, keyword opportunities, index coverage, technical SEO issues and trends.
• User funnel shows audience over time, new users and key events, and retention by cohort.
• Reliability shows repository health.
• Timelines read oldest to newest, left to right.
• A table nobody has written to yet says so, and a value nobody has measured is never shown as zero.

Models

Local models see how much memory your Mac really has

OpenHelm now reads available memory the way macOS reports it, so a Mac with half its memory free is no longer treated as nearly full.

• Local runs can use a larger context window when the memory is there.
• The window is sized for the whole run, not just the opening prompt.
• Jobs load only the tools their task names, leaving more room for the work.
• Job runs use a steadier sampling setting than chat.

Desktopv2.3.1

Desktop sign-in fixed

Earlier desktop builds could not sign in to OpenHelm Cloud or reset a password. 2.3.1 ships with the correct cloud connection built in.

Desktopv2.3.0

OpenHelm 2.3 for Mac

The first desktop release with local Gemma models, plus a fix for settings that belong to one Mac rather than your whole account.

• Choosing which AI backend a Mac uses now reaches that Mac, instead of being saved to the account and never applied.
• Each Mac confirms when it has applied a change.
• The same per-Mac settings can be managed through the API and MCP tools.

Grow

Pillar and cluster linking, and a monthly keyword check

Your keyword plan now records which pillar each cluster topic supports, and posts on both sides are linked to each other as they are published.

• The strategist names the pillar keyword behind every cluster keyword.
• The publisher adds internal links both ways between pillar and cluster posts.
• A monthly keyword accuracy review checks each post still matches its target keyword.
• Clear drift is corrected in metadata only, with a cap per run. Cannibalisation and anything unclear go to you for review.

Grow

Sitemaps are resubmitted once a fix lands

Google offers no public API for requesting indexing of a single page, so technical SEO agents now resubmit your sitemap through Search Console instead, which is the documented route.

• The sitemap freshness check resubmits rather than only flagging.
• The index coverage job resubmits once it confirms a fix is live.
• Nothing drives Google’s own interface behind your back.

Grow

Trend and competitor research join the product crew

The product template now includes a trend researcher that scans your niche each week and a competitor analyst that writes a weekly brief.

• Trends are recorded for the blog publisher to draw on, while your keyword plan stays in charge.
• Both use built-in web search, so they need no extra connection.
• New products get both agents straight away. Existing products are offered them for your approval.

Models

Run agents on Gemma, entirely on your Mac

A third backend beside Claude Code and Codex: Gemma 4 served by Ollama, for jobs, chat and planning with no subscription at all.

• Installs Ollama for you through Homebrew where it can, and says so plainly when you need to finish with the official installer.
• Pick a model size with memory guidance and a live download bar.
• File access is kept to the job’s own folder.
• Experimental: smaller models handle focused jobs better than long multi-step ones.

Understand

Analytics set up for you

When a live product has no analytics, the task OpenHelm files can now do the setup: create a GA4 property and web data stream, link them to the product and add the measurement ID to your deployment.

• Uses your own connected Google Analytics account.
• Stops and asks if more than one Google Analytics connection could apply.
• Creating a brand new Google Analytics account still needs you to accept Google’s terms.

Fix

Sentry issues reach your agents through Contextely

Sentry issues relayed through Contextely now wake the right product’s incident job. We tested it end to end with a real issue.

• Issues are routed to a product by matching the Sentry organisation and project names you have mapped.
• Each hop is signed and checked by whoever holds that key.
• A feed that stays silent after setup now files a task instead of looking healthy.

Fix

Bug fixes go from report to resolved

The jobs that act on a Sentry issue or a customer’s bug report now check the request first, then fix, test, deploy, verify and close the loop with whoever reported it.

• A legitimacy check asks: is the report real, in good faith, and consistent with decisions already made?
• Bad-faith requests are refused and escalated, never partly carried out.
• The fix runs your project’s own tests and adds one where a test could have caught the bug.
• An issue is resolved only once the fix is verified live. Merged but not shipped leaves it open with a comment.

Platform

Five priority levels for jobs

Every job now has a priority from Ultra low to Ultra high, and new jobs must pick one.

• Higher priority gives a head start in the queue, and older runs still get through.
• Work somebody is waiting on, like a customer email or a live incident, is Ultra high by default.
• At low spend settings, the lowest-priority recurring jobs pause first. Manual runs are never held back.

Platform

Webhook setup you can finish inside the app

Any connection that can push events to OpenHelm now has a setup panel showing its webhook address, the steps to follow and a place to save the signing secret.

• It only says it is receiving once a real delivery has arrived.
• Webhook is now a trigger you can pick when creating a job.

Platform

Plugins: ThreadCamp for mail, Contextely for connections and memory

OpenHelm now has plugins. ThreadCamp provides Mail, and Contextely provides Connections and Memory.

• Until you add a plugin, its screen shows what it would switch on and loads nothing.
• Where Contextely can connect a service, the connection form hands off to it.

Platform

Cloud runs remember

Memory now lives in Contextely, and cloud runs carry the product’s memory into their prompt for the first time, as desktop runs already did.

• Memory is scoped to one product and never read across products.
• Memory has its own screen next to Mail and Connections.

Reply

A customer’s reply wakes the job that answered them

When a customer replies "that didn’t work", the product’s triage job now runs again, so the support conversation can continue.

• A product’s inbox is created linked to that product.
• Mail sent from any of your own inboxes never wakes a job, so two products cannot answer each other in a loop.

Reply

The hello@ address on your website receives

The address your site actually publishes, hello@ on your own domain, can now be a real inbox, where before mail to it could be dropped.

• OpenHelm only takes that address when it can see your domain’s mail is not already handled elsewhere.
• If it cannot tell, it asks rather than guessing.

Grow

Every post carries its own Search Console numbers

The weekly SEO digest now writes clicks, impressions, click-through rate and average position back to each published post, so you can see which posts earn and which to prune.

• Each figure is dated, so "measured and nobody saw it" is never confused with "not measured yet".
• A failed Search Console pull writes nothing rather than re-dating old numbers.
• The strategist keeps the Keyword Catalogue in step with the content plan.

Reply

Desktop answers product mail, and one complaint no longer silences you

The desktop app now checks your products’ real inboxes and wakes the matching job, just as the cloud does.

• A single spam complaint no longer pauses an inbox. It takes several people, or a complaint rate above 0.5% once you have sent at least 200 messages.
• If an inbox is paused, OpenHelm files a task so you know it cannot send.

Reply

Inbound mail reliably wakes a job

Several fixes so customer email always reaches an agent, including mail sent to forwarding addresses.

• Mail to a forwarded address is still routed to your agents, not only copied to a person.
• Mail that arrives while your Mac is asleep waits and is retried, rather than being written off.
• Each message in a thread shows whether it has been routed, and to which job.

Reply

Customer replies follow your risk setting

Your approval risk level now decides when a routine reply sends by itself and when a draft waits for you in Tasks.

• Approve and Reject in Tasks send or discard the held reply.
• Replies stay on the customer’s original thread.
• Mail now runs on ThreadCamp in both cloud and desktop.

Understand

User journeys and funnels for each product

A journey analyst now studies each product’s code, routes, pricing and analytics before defining its key user journeys, then tracks them from Google Analytics funnel reports.

• The User funnel screen shows each funnel, likely stuck points and the tasks filed about them.
• Missing step data is shown as missing, never as zero.
• The analyst’s jobs start switched off until you approve them.

Reply

Mail follows the product you are looking at

Pick a product and Mail shows only its addresses, conversations, approvals and forwarding. All products remains one click away.

• Conversations open alongside the list, so replying never loses your place.
• A calmer visual design across the whole app.

Models

Run on your Codex subscription

Desktop users can now choose Codex or Claude Code as the AI that runs their agents, chat and planning.

• Your Codex usage and remaining capacity show in the app.
• If the chosen backend is missing or signed out, OpenHelm tells you rather than switching quietly.
• OpenHelm never reads or stores your Codex credentials.

Reply

Mail forwarding

Forward one address or everything at your domain to another inbox, such as Gmail, with forwarding set up to keep deliverability intact.

• Bounces are handed back to the original sender.
• Loop protection is built in.

Reply

OpenHelm Mail replaces your transactional provider

Products can send all their mail through OpenHelm Mail, including password resets from tools that only speak SMTP.

• SMTP sending with authentication, using your OpenHelm API key.
• Transactional mail and agent mail have separate daily limits, so a campaign can never block a password reset.
• Reply-To only points at an address that can actually receive.
• Cc and bcc supported, with suppressed addresses dropped and reported.

Desktop

Your Mac can run your cloud jobs

Signed in to the desktop app, your Mac picks up cloud jobs and runs them on your own Claude Code subscription, using no AI credits.

• Choose per job whether it runs on your Mac or in the cloud.
• Choose what happens when your Mac is offline: wait for it, or run in the cloud.

Reply

Real inboxes for your agents

OpenHelm Mail gives agents their own inboxes with threads, approvals and custom domains, all visible in a new Mail tab.

• Each inbox sends automatically, sends and notifies you, or holds messages for approval.
• Bring your own domain. OpenHelm generates the DKIM records to add.
• Bounces and complaints are suppressed automatically.

Grow

Organic growth agents

A set of agents that plan keywords from live search data, pace publishing to what Google is indexing, prune what stops earning and make your site readable to AI answer engines.

• Keyword research from Google Ads Keyword Planner data.
• A weekly content budget set from indexing health, capped at three posts a week.
• Content lifecycle passes that refresh, merge or prune posts using Search Console evidence.
• Honest link building from directories, resource pages and personal outreach.
• GEO audits for llms.txt, structured data, AI crawler access and citable facts.

Growv2.2.0

Site and competitor monitoring

Agents that watch a site daily for meaningful changes, send alerts with evidence and write a weekly source-cited competitor brief.

• "No meaningful change" is a proper result, not a failure.
• Pages that cannot be reached are reported as such, never guessed.
• Desktop runs now recover cleanly from sleep and from hitting a usage limit.

Platformv2.1.0

A new Autopilot engine

Every agent now has goals it has to prove, a separate evaluator judges each run, and hard daily budgets cap what Autopilot can spend.

• Each agent carries two to five checkable goal criteria. Progress means criteria met, not activity.
• A fresh evaluator scores every run against its outcome contract and leaves a hint for the next one.
• A daily pass tidies memory and surfaces recurring failures.
• When nothing has changed, a check costs no tokens at all.
• Works the same on desktop and in the cloud.

Platformv2.1.0

Chat that actually sets things up

Describe an outcome in chat and it creates the agents, jobs and tables for real, either straight away or as one card you approve.

• At full autonomy changes apply immediately. Otherwise they wait on a single approve-all card.
• Every job created in chat carries an outcome contract: what done looks like, how to check it and when to stop.
• If chat promises something and does not create it, OpenHelm catches it and follows through.

Platformv1.3.0

A spend dial, and Autopilot that acts on failures

One five-step dial now controls how much OpenHelm spends, and cloud Autopilot steps in when a job keeps failing.

• Lower settings run recurring jobs less often on cheaper models. Higher settings do the opposite.
• After three failures in a row, Autopilot pauses the job or asks you first, depending on your risk setting.
• Metrics that are off track after a run become tasks.
• Reconnect is always available on a connection, even before its token is known to have expired.

Platformv1.2.0

Team workspaces

Bring your team into OpenHelm Cloud with shared workspaces, roles and billing at the organisation level.

• Owner, Admin, Member and Viewer roles.
• Invite teammates by email.
• Jobs can now run after another job finishes, or when an email arrives at your organisation’s address.

Platform

A public API and MCP servers

Developers can trigger runs and manage connections through the REST API, or use OpenHelm from any MCP client.

• MCP servers for research, SEO, email outreach and code changes.
• A Connections API with hosted connect links.

Platform

Tasks and approvals

Anything an agent wants you to decide or sign off now arrives as a task with clear actions, in one place.

Platform

Goals become agents

Goals are now agents, each with a persona, its own chat thread and memory, and the ability to hand work to sub-agents.

Desktopv1.0.0

Voice chat on your Mac

Talk to OpenHelm out loud. Speech recognition and speech output both run on your Mac.

• Conversation mode, or push to talk.

Platformv0.5.0

Charts from your data

Line, bar, area, pie and stat charts drawn from your data tables, and agents can create them during a run.

• Several chat threads per project.

Platformv0.4.0

Data tables your agents read and write

Structured tables that you and your agents both edit, with eight column types and a history of every change.

• Every change records whether you, an agent or the system made it.

Desktopv0.3.0

A built-in browser for agents

Agents get their own browser with no setup, so jobs can work on websites that have no API.

Platformv0.2.0

Autopilot

Describe a goal and OpenHelm drafts a plan of scheduled jobs. Nothing runs until you approve it.

• A secure store for the API keys and tokens your jobs need.

Desktopv0.1.0

OpenHelm for Mac

The first release: schedule jobs that run in the background on your own Claude Code subscription.

• Watch each job’s output live as it runs.
• Failed jobs retry with what went wrong added to the prompt.
• Runs interrupted by a crash are recovered on the next launch.