Changelog

August 29, 2026v1.10.0

X-rays

The way work actually gets done in your organization is already encoded in the traces your agents produce. An X-ray is one question asked of that record, answered across a scope you choose.

Describe summarizes a scoped set of traces. Workflows finds the procedures being run over and over, including the ones different teams arrive at through different tools, and sketches the tool that would collapse each into a single call. Wiki builds a documentation corpus out of what it observes, writing up the workflows, subsystems, tools and practices it finds, each page cited back to the sessions behind it and never to work the run did not read.

Each new X-ray asks a new question of traces you own.

Improvements
EndpointsCensus readings group every session into the organization-wide initiatives it belongs to, and each session shows which ones it joined.
EndpointsA Census tab in the generation picker lists your readings and starts a new one.
EndpointsInitiatives show their lifespan and can be compared by the evidence they share.
EndpointsAdmin and AI governance have their own places in the navigation, separate from personal settings.
EndpointsLanes shows the trace reading happen call by call, and marks what it confirmed, disputed, and was still looking at.
EndpointsThe Lanes timeline draws on a canvas, so it stays responsive across a whole session.
EndpointsThe joins dock finds candidate hand-offs through a content fingerprint index.
EndpointsMicrosoft 365 Copilot traces carry a chip for the Office app they came from, and reasoning is disclosed inline.
DetectionSignal findings carry typed evidence spans, so what a trace highlights is what the detection decided.
DetectionCredential access is detected as a read or a write, naming the resource it reached.
DetectionPrivilege escalation in agent commands is detected.
DetectionAgent edits to AI harness configuration files are detected, available per organization.
DetectionUnicode smuggling detection was rebuilt to require a real steganographic alphabet.
DetectionCredential detection drops credential-shaped strings that are not credentials.
DetectionAI providers you add yourself join the catalog agents watch, alongside the sanctioned and unsanctioned classification.
IT & rolloutClaude Code running inside WSL is captured, with its policy provisioned in the Linux environment.
IT & rolloutReinstalling a machine reuses its endpoint identity instead of consuming another quota slot.
IT & rolloutMicrosoft 365 Copilot is detected inside Word, Excel, and PowerPoint, not only as a standalone install.
IntegrationsGoogle Workspace connects and scopes from the dashboard, with related people, groups, and devices linked from each details flyout.
IntegrationsGoogle Workspace catches up incrementally between full syncs.
IntegrationsEvery directory provider shares one setup flow.
IntegrationsSignal detections can be delivered to Slack instead of waiting for someone to open the dashboard.
Fixes
Google directory sync keeps whole-domain group membership across catch-up syncs, and retires memberships from the group side for every provider.
The Google connection test sends the parameters it needs.
Microsoft 365 connector fetches recover from dropped pages, and spreadsheet responses are read correctly.
Microsoft 365 Copilot traces show only turns that carry real data.
macOS agent upgrades are bounded and recover instead of wedging.
Cowork transcript progress survives a restart.
Prompts are spooled durably through backend outages instead of being lost.
Codex code-mode calls are decoded instead of left wrapped.
Re-registered endpoints appear once in the sidebar, and report the hostname they registered with.
Signal highlights land on the right span for every active signal.
Credential evidence spans resolve when the surrounding text was trimmed.
A search that reads no file content no longer counts as credential access.
A version flag with no operand reads as a query rather than a package install.
Links from the security posture canvas route to the right detection.
A canvas publishes the kernels it depends on.
Analytics queries keep their LIKE escapes, and string parameters are escaped.
An over-scanned query offers the narrower window again.
The Lanes timeline repaints when the theme changes.
Agent surfaces are labeled by provider when the agent type is not reported.
August 22, 2026v1.9.12

Lanes

Multi-agent systems are becoming a new kind of computer. Their execution graph is written at runtime. Agents that never share a context window still pass work through files, tool results, commits, and handoffs. The transcript records the messages and loses the machine.

Lanes makes that machine visible. It reconstructs the run as a map across agents and artifacts, with provenance showing how earlier work shaped what happened downstream. You can see the system fork and rejoin, then open any point to inspect the exact exchange or file state.

Agent swarms are becoming infrastructure. Lanes makes them observable.

Lanes maps the work of an agent run

Local inference, observed

Ollama, LM Studio, vLLM, and software built on PyTorch or MLX can run models without a request ever reaching a cloud provider or AI gateway. Provider and gateway logs have nothing to record.

Origin sees the work where it runs. The endpoint records when a known runtime starts and when a process loads a recognized inference library, even inside a custom application with no server port or network connection. Accelerator-backed loads are reported separately as stronger evidence that the process ran inference on the GPU.

Every observation stays tied to the endpoint and process that produced it. Inventory separately shows installed local runtimes and, when available, their versions. Security teams can distinguish software that is merely present from endpoint evidence of execution.

Improvements
EndpointsTrace Review runs an agent over a single trace and returns a verdict ledger, with dissenting reads surfaced rather than averaged away.
EndpointsWiki pages assemble themselves from a scope's agent work and keep themselves current, one page set per scope.
EndpointsA Topic patterns lens reads which subjects turn up together across sessions, and follows the scope you have steered to.
EndpointsThe Discover tab scouts a scope for methods worth reading, and opens any of them as a trace.
EndpointsCanvases can be arranged in place: drag tiles to order the rows, drag the dividers to set widths.
EndpointsCanvases and kernels are stored in your account, so they follow you between browsers and machines.
EndpointsNew tenants get pre-built canvases tailored from a profiling run of their own activity.
EndpointsA security posture canvas puts the fleet's AI picture in one view for a security lead.
EndpointsTrace detail leads with an anatomy view of the run, and shows the request's effort setting and whether fast mode was on.
EndpointsCodex code-mode cells and attachment results render as what they are, instead of as opaque blocks.
EndpointsEndpoint Inventory filters and sorts server-side, with a live row-range label as you scroll the fleet.
EndpointsInstalled AI tools carry their product brand marks in Inventory.
EndpointsInventory finds self-hosted serving stacks and reads their versions from disk, including tools installed through Python and npm package managers.
EndpointsSearch a trace by a person's name, email, or email domain.
EndpointsDashboard chat conversations are saved and restored by default, so a session survives a reload.
EndpointsA brand-new tenant lands on a Connect front door that walks it to a first reporting endpoint, with empty states that say what is still missing.
EndpointsGrok Build is captured as a supported agent, with prompts, responses, tool calls, and per-turn token usage.
EndpointsExplore is the canonical view for LiteLLM gateway activity.
EndpointsThe investigation surfaces that used to sit behind an experimental flag are on by default, with their own permission and availability rules unchanged.
DetectionA package-install signal reports what agents add to a machine, decided by the shape of the shell command rather than a substring match.
DetectionA file-activity signal records what agents read and wrote, with authorship, replay order, and Codex patch applies.
DetectionCatastrophic-command detection was rewritten, closing additional false-positive classes.
DetectionA signal flags when an endpoint's managed settings keys are shadowed or overwritten.
DetectionCredential detection reads the command line as well as file contents.
DetectionSignals can be scoped to selected endpoints, and the dashboard reads the live signal catalog from the backend.
IT & rolloutOrigin runs a production environment in Australia for organizations that need their data to stay in region.
IT & rolloutThe macOS installer raises itself when it finishes, recovers when full-disk access is granted mid-install, and accepts paste in the registration dialog.
IT & rolloutThe tray reports its version as soon as it starts, before the agent checks in.
IT & rolloutOrigin support can be granted read-only access to a dashboard to work a ticket, and any organization can switch that access off.
IntegrationsOkta connects through an Origin-managed service app, with group-scoped sync, incremental updates, and computed member counts.
IntegrationsGoogle Workspace supports scoped group sync and device sync.
IntegrationsMicrosoft Entra picks up membership changes between full syncs, and writes directory records as they stream in.
IntegrationsAny directory provider can be disabled and re-enabled without deleting the integration.
IntegrationsAfter a directory sync, activity that had no owner is swept up and attributed.
IntegrationsChatGPT conversations can be brought into the same prompt pipeline as the rest of your captured AI work, enabled per organization on request.
IntegrationsPrompt resource paths are a filterable dimension, and per-request latency is available on the token entity.
IntegrationsThe analytics graph can be navigated along conversation lineage.
Fixes
Explore Inventory loads the whole fleet instead of stopping at the first page.
An AI tool detected through several paths shows as one row, not several identical ones.
Cost estimates price current models, include cached input, and say when their coverage is partial.
Records that are not user conversations stay out of prompt lists.
Gateway-captured content appears in the Explore list.
Prompts are no longer lost when a message is cancelled in flight or arrives coalesced.
A prompt with a large attachment is no longer dropped for size.
Codex prompts, custom tool calls, and shell tool records are captured across recent versions, with stable turn identity and sub-agent lineage.
Shell file activity is attributed to the agent that ran it, and holds across turn drift.
A failed shell write says whose failure it was.
Suspending a machine is detected on a clock that counts sleep.
After a burst, the agent stops re-sending bytes it has already uploaded.
The macOS tray launches correctly when more than one user is signed in.
Uninstalling on macOS removes the tray agent.
Shell-invoked hooks run on signed macOS builds.
Endpoint-reported state survives re-registration.
Configuration changes on an endpoint are attributed to whoever made them, with Origin's own writes distinguished from everyone else's.
Removing a directory integration deletes it durably, cleans up managed consent, and can be re-authorized afterwards.
A directory integration keeps its scope and its status when it is updated or refetched.
One sync runs per directory integration at a time, and an aborted run recovers on the next cycle.
Entra group previews retry while permissions are still propagating.
A removed Google group no longer retires the whole directory.
Group scope is reviewable before you save it, the save syncs, and errors say what went wrong.
A pending invitation is resent instead of failing when the person is invited again.
Microsoft 365 Copilot sync tracks progress per user, so one slow account does not hold up the rest.
Microsoft 365 Copilot answers render from card attachments instead of arriving empty.
Trace queries are bounded, and a scan rejection names the filter that will bring the query into range.
The Files view keeps its scroll position when you move through Prompts.
Related-work traces open directly, and signal origin and span anchors survive the jump.
Long paths, tooltips, canvas headers, and onboarding steps hold their layout at narrow widths.
Chat restore brings back images and attachments, groups submissions correctly, and attributes sub-agents.
Canvas divider widths persist, and the security posture canvas stops re-running its kernels when you switch chats.
Actions are hidden on your own row in user management, and its column picker counts correctly.
The invite flow and password copy use your organization's language.
Every role can attach a file to a bug report.
The MCP setup dialog scrolls instead of overflowing.
Windows capture bounds the buffers a monitored process can cause it to allocate.
Removing a skill can no longer be steered into deleting files outside that skill.
An event Origin cannot convert no longer fails the whole page of results.
Dimension acronyms keep their capitalization in copy.
August 4, 2026v1.8.5

Governance you can verify

Putting AI agents to work across a company means deciding the terms they operate under. Origin sets those terms on the endpoint itself, so an agent works the way you decided, and what it does becomes a record you can review.

Policy can be displaced, though. Another system in your environment can set its own terms for the same agent, and from that point yours are not the ones in effect. Nothing about it looks wrong. The machine stays enrolled and healthy, the agent keeps working, and the record coming back from it quietly thins out.

Origin now watches the policy it set, and tells you when something has taken precedence over it or changed what it depends on. The finding stands for as long as the condition does. Governance you cannot verify is an assumption, and this is the part that makes it a fact.

Available per organization, on macOS and Windows.

Shared canvases

Until now a canvas lived in the browser that built it. You could assemble exactly the view your team needed for Monday and be the only person who ever saw it.

Canvases and the kernels behind them are now stored with your organization, so building one makes it available to everyone you work with. They follow the drilldown: open the same canvas on a single endpoint or one person and every widget re-runs narrowed to what is in front of you. Forking a built-in canvas gives the team something to work from rather than a private copy.

Editing changed too. Drag cards to reorder them and group them into rows, drag the divider between two cells to set how much width each one takes. What you arrange is what renders, in view mode and in exports.

Improvements
IT & rolloutA new Australian deployment region has been added
IT & rolloutThe macOS agent no longer installs a network extension, and the configuration profiles that existed to provision it are retired.
IT & rolloutParsers for each captured format are now released on their own, so support for a new agent or format reaches endpoints without an agent update.
DetectionOrigin reports when the policy it set on an endpoint has been displaced, or when what that policy depends on is changed by something else.
DetectionPackage installs run by agents can be enabled as a signal per organization.
DetectionSignals can read endpoint events, not just prompts.
DetectionSignals that describe an endpoint rather than a session now have a place to be read in the dashboard.
DetectionRetiring a signal removes it everywhere at once, instead of leaving it running out of view.
EndpointsEndpoint Inventory filtering and ordering run server-side, so sorting and filters cover the whole scope instead of the rows the browser happened to load.
EndpointsInventory shows which rows you are looking at, and says when rows were merged by hardware identity.
EndpointsTrace search finds people by name, email, or email domain, including Copilot and API accounts.
EndpointsTrace detail shows the effort and speed an agent asked for, alongside model, tokens, and tools.
EndpointsPer-request latency reads alongside token usage in analytics.
EndpointsCost estimates price current models, state what the estimate is based on, and hold back when pricing coverage is incomplete.
IntegrationsDirectory sync re-sweeps activity it could not attribute earlier, so a person's record fills in retroactively instead of staying blank.
IntegrationsThe dashboard assistant runs on Origin-managed inference by default, with no gateway to connect.
DesktopThe tray shows the installed version right away, before the agent reports in.
DesktopFile activity capture skips build output and telemetry spool directories, picks up dependency lockfiles, and reports renamed files under their final name.
Fixes
Native dropdowns stay readable in dark mode.
Every role can attach a file to a bug report.
Inviting someone who already has an invitation pending resends it instead of failing.
The user management column picker counts columns correctly, and its actions column is labeled.
The Observed Skills origin filter returns results instead of an error.
Analytics time ranges bucket correctly again.
Directory sync runs one sync at a time per integration and recovers from an interrupted run instead of stalling.
Removing a directory integration finishes more reliably, cleans up managed consent, and integrations can be re-authorized.
Microsoft Entra group previews retry while permissions are still propagating.
Group scope shows what is actually reviewable, saves cleanly, and reports errors in plain language.
Jumping to a signal firing stays anchored in trace detail once highlighting finishes.
The MCP tab scrolls inside the setup dialog instead of overflowing it.
Canvas exports flag fewer ordinary identifiers as credentials.
Prompt content is no longer written to service logs, which reference the stored record instead.
Resumed Codex sessions attribute their tokens correctly, and their effort telemetry is preserved.
Endpoint configuration changes keep the attribution of who wrote them.
The macOS installer detects Full Disk Access reliably, returns to the front after setup, and accepts paste in the registration dialog.
The macOS installer package no longer carries an architecture suffix in its filename.
Uninstalling on macOS removes the tray helper, and the tray launches correctly when several users are signed in at once.
Removing a skill can no longer delete files outside its own directory.
Staged Store updates on Windows pick the newest package.
The agent keeps capture running when the host is under memory pressure, and only warns about backend reconnects when they persist.
Microsoft 365 Copilot webhook subscriptions handle contention and scope matching correctly.
Related work opens its trace directly.
Studio tabs and pinned items restore correctly, canvas column widths persist, and the inventory table scrolls as rows arrive.
July 17, 2026v1.7.4

The people behind the prompts

A hostname is not a person. As AI spreads across a company, understanding the work means knowing who it belongs to and how that person fits into the organization.

Directory integrations are now generally available for Microsoft Entra ID, Google Workspace, and Okta. Connect the directory you already maintain and Origin brings its people, groups, and organization details into the graph. When an endpoint can be matched to its directory owner, Explore rolls that person's activity together across devices, with their department and email alongside it.

All three integrations can sync a full directory. Microsoft Entra can instead start with selected groups, including nested membership, so a focused rollout does not require bringing in the whole organization.

Traces attributed to a person

Microsoft 365 Copilot, in context

AI adoption does not stop at developer tools. With Microsoft 365 Copilot, the same people are working with AI in Word, Excel, Outlook, Teams, and Copilot Chat.

Enable capture on a connected Microsoft Entra directory, and Origin brings those conversations into the same activity view as the rest of your captured AI work, attributed to their authors. The same review workflow now covers coding agents and the work happening across Microsoft 365.

Microsoft 365 Copilot activity in a trace

Improvements
IntegrationsExpired directory credentials can be updated in place from the integration's settings, without deleting and re-adding it.
IntegrationsMicrosoft Entra directory integrations can sync selected groups, including nested membership, or the whole directory.
IntegrationsOrigin's MCP server can search prior work and finds similar past sessions even when the wording differs.
IntegrationsOrigin's MCP endpoint now uses streamable HTTP transport, which more clients support.
IntegrationsThe dashboard assistant can use Origin-managed inference, enabled on request, so you don't have to connect your own gateway.
EndpointsEndpoint last-seen times now come from agent heartbeats.
EndpointsEndpoint Inventory search runs server-side across the whole fleet.
EndpointsInventory detects AI extensions across VS Code forks, including VSCodium, Insiders, and Windsurf.
EndpointsInstalled agents are grouped under friendly product names, and search matches the names you see.
EndpointsCanvases arrange tiles in a flexible grid with cleaner tile chrome.
EndpointsTables and heatmaps handle wide results better, and PNG exports can span the full width.
EndpointsYou can launch a steered clustering pass from Explore's generation picker.
EndpointsExplore's Memories dimension is now driven by the memory signal.
EndpointsWeekly prompts-per-user distributions report levels beyond 1,000.
DetectionSignal firings are highlighted in place in trace detail, with tooltips and a whole-trace scan.
DetectionSignals beyond the built-in baseline can be enabled per organization.
IT & rolloutCapture and detection updates can now roll out to endpoints between agent releases.
Fixes
Chat is hidden for roles that don't have permission to use it.
The skill signal now catches slash-command invocations.
Malformed analytics time ranges return a clear error instead of failing the query.
Treemap values stay legible over cell patterns, and cells show tooltips.
Every install path now configures the agent's backend connection.
AI tools with spaces in their executable paths are detected more robustly.
Tray diagnostics only flag issues for capture paths enabled on that endpoint.
Trace and command deep links open in their page context instead of the chat panel.
Captured activity is labeled with its true source instead of showing as proxy capture.
Prompts that arrive out of order are recovered more robustly.
Very long sessions and agent reconnects no longer truncate traces, and the current turn always renders.
Captured events survive agent reconnects to the backend.
Agent-harness scaffolding and tool output are no longer misclassified as prompts.
Signal-scoped token counts are no longer under-reported.
Microsoft Entra directory sync no longer stalls on empty pages or spawns duplicate runs, and large groups sync fully with scope changes applying within one cycle.
Removing a directory integration shows its status correctly until it finishes.
Microsoft 365 Copilot content renders fully instead of showing unknown blocks.
The dashboard assistant no longer errors on models that require extended thinking, and model errors surface instead of failing silently.
Chat's save suggestion appears after the response instead of interrupting it.
Dashboards load for organizations whose names contain parentheses.
Credential detection is consistent between browser and endpoint capture.
The MCP server handles JSON-RPC notifications correctly.
Installed AI tools no longer appear twice when detected through multiple paths.
July 2, 2026v1.6.7

Signals

Signals are deterministic rules that annotate the graph. Each looks for a specific, well-understood pattern in prompt text, tool calls, tool results, or tool definitions, and marks the exact activity where it matches. A signal is a finding, not a guess: it fires because a rule matched, not because a model thought something looked off.

The first set we’re publishing is security detections, on by default and maintained by Origin: credential and financial data, prompt injection, risky commands and exfiltration, tool poisoning. Each firing is anchored to the activity it found, so you can jump straight from a finding to the surrounding trace, or read it as a dimension across the fleet in Explore.

Signals firing in trace detail

Role-based access control

Not everyone who uses Origin needs the same view of it. Until now, everyone who wasn’t an admin collapsed into a single "Member" role, so access was effectively all or nothing.

Role-based access control replaces that with a set of built-in roles that allow for more granular control, assigned by admins from User Management. Enforcement runs across both the interface and the underlying data: each role sees only the sections it should, analytics queries are scoped to the endpoints and data it’s allowed to read, and prompt content stays gated even for users who can see that a signal fired. Roles can map to your directory groups, so access follows the structure you already maintain.

Improvements
IntegrationsSync selected groups from your identity provider so people and devices become first-class dimensions in analytics, with support for Microsoft Entra, Google, and Okta.
IntegrationsMicrosoft Entra can be connected through Origin-managed admin-consent onboarding, with no app to register or maintain yourself.
IntegrationsPoint the dashboard assistant at your own OpenAI-compatible LLM gateway with a custom URL and token, governed by a per-organization host allowlist.
IntegrationsOrigin's MCP server is now available at a stable first-class endpoint at mcp.originhq.com.
IntegrationsDirectory sync surfaces each user's country from your identity provider as a jurisdiction-attribution source.
EndpointsExplore has a new structured query bar: type filters with autocomplete and editable pills, and click a chip to select its range in the query.
EndpointsSearch clusters and endpoints by label in Explore, and refine any sidebar breakdown in place with a layered filter.
EndpointsEstimated cost now appears alongside token usage across charts, footers, and trace details.
EndpointsA new dashboard home brings canvases and chat together as a single front door, with pre-built read-only kernels and canvases to start from.
EndpointsEndpoint inventory adds a fleet-composition view with agent and OS mix, install footprint, and first-seen and last-seen trends.
EndpointsInventory dedupes duplicate and re-registered endpoints, including VM clones, by hardware identity.
EndpointsTool and MCP-server detail pages show their definitions, how they drift over time, and how they are being called.
EndpointsExplore adds a Memories dimension showing the agent memory each person's sessions touch.
EndpointsCodex activity is now captured with per-turn tokens and the tools it runs.
EndpointsSessions are labeled by their most recent prompt summary instead of a raw identifier.
EndpointsThe trace sidebar shows where a session originated: the launching agent, the user, and the command.
EndpointsTrace detail captures more session context from the full request, including working directory, git branch, and platform.
EndpointsNetwork, DNS, file, and process-spawn events are deduplicated per process and session.
IT & rolloutUninstalling the agent captures a diagnostics bundle and files it automatically.
Fixes
Expired assistant tokens now return a clear error and refresh proactively instead of an empty response.
Web-based AI prompts show the real model instead of "unknown".
Prompt cards now render for claude.ai, Claude Desktop, and DeepSeek web.
Chat tables render with clean headers, working tooltips, and matching PNG exports.
The activity surface now sizes correctly in Safari.
The dashboard keeps your saved theme on the sign-in screen instead of flashing the default.
Added a recovery path out of a wedged dashboard sign-in.
Hardened attachment serving against a stored cross-site scripting risk.
Analytics results resolve endpoint IDs to readable names in bars, rows, and timelines.
Trace log entries no longer get stuck on "Pending": harness scaffolding reads "Infrastructure", promptless terminal steps are labeled, and outlier traces show their cluster badge.
The analytics assistant now answers cluster-scoped token questions instead of dead-ending.
Endpoint inventory closes AI-tool detection gaps on Windows and macOS.
Endpoint inventory now shows the macOS version.
macOS capture keeps working under encrypted DNS and iCloud Private Relay.
Prompt and usage capture now reads Brotli-compressed responses.
Microsoft 365 Copilot capture no longer duplicates prompts from control frames.
Claude.ai telemetry beacons and ChatGPT prewarm requests are no longer captured as prompts.
macOS capture self-heals if the network extension's flow limit saturates, and recovers from filter revert loops.
Windows installs honor the selected drive and install directory.
The macOS installer reports a clear error for a malformed enrollment token instead of failing silently.
The agent recovers cleanly from watchdog restarts, avoids crash-loops when the network is unreachable at boot, and no longer rolls updates back during a coordinated restart.
Directory sync is more resilient to transient identity-provider errors across Microsoft Entra, Google, and Okta.
Improved coverage of built-in credential and secret detection, restoring long-tail detectors and closing false negatives.
June 12, 2026v1.5.15

The proxy-less agent

Version 1.5.15 introduces a proxy-less agent for collecting AI prompt activity from endpoints without placing anything in the network path.

The agent captures supported prompt activity directly from AI tools on the machine and ties it to the person, device, and surrounding endpoint activity. Prompts land in the same context graph as processes, files, and network connections, and show up in the same traces, analytics, and inventory, so teams can investigate AI-assisted work alongside the rest of the endpoint record.

Proxy-less collection requires no traffic interception, certificate changes, or new network trust. Install the agent and supported activity starts flowing into Origin. For organizations starting their AI visibility journey, this is the fastest path to a live record of AI work, and everything it captures carries forward as endpoints later move to the full agent.

The full agent remains the deepest view of an endpoint, observing activity on the wire in addition to endpoint context. The proxy-less agent is designed for faster deployment and coverage of tools that report their own activity.

Both paths build the same graph of AI work in Origin.

Improvements
EndpointsPrompt activity from Claude Cowork sessions is now captured on both macOS and Windows.
EndpointsClaude's browser-based agents now show up in activity, including the tools they execute.
EndpointsProxy-less collection drops model response bodies on the endpoint instead of uploading them.
EndpointsCapture recovers automatically if other software stops the agent's collection session.
EndpointsFleet health reads proxy-less endpoints correctly: the network capture component shows as not applicable instead of unhealthy.
EndpointsThe agent adoption chart stays readable at fleet scale, and endpoint detail now shows version skew and stability.
EndpointsEndpoint details now omre robustly distinguish Windows 11 from Windows 10.
EndpointsToken usage is attributed at a finer grain, includes provider-reported cache tokens, and breaks down into input, output, cache, and reasoning tokens across charts and trace details.
EndpointsPrompt usage can be broken down by the client surface it came from.
EndpointsTrace detail adds Context, Tree, and Raw views and now leads with the prompt summary and endpoint.
EndpointsExplore adds a Skills dimension showing activity by the skills agents used.
EndpointsExplore can pin a clustering generation and drill through cluster hierarchies.
EndpointsFlow links in Explore connect model and cluster scopes, with hover pop-outs and drag-to-pan.
EndpointsExplore's Context and Chat panels moved to a right-edge mode rail.
EndpointsEndpoint and session chips in event detail are clickable and jump to that scope.
EndpointsThe dashboard assistant runs on Claude's newest models and can open Explore scopes and Studio items for you.
EndpointsDashboard chat adds rewind and fork, markdown rendering, and kernel save suggestions.
EndpointsSigning out clears Origin's local browser data, and Settings can clear it on demand.
EndpointsDashboard session tokens are no longer persisted in browser storage.
EndpointsCleaner dashboard URLs with no # in the address bar, direct settings links, and shared links that survive the sign-in redirect.
EndpointsLarge datasets render faster across the dashboard: chats lazy-load, long trace views virtualize, and canvases coalesce redraws.
DesktopRun Diagnostics in the endpoint tray shows the agent's service identity, snapshot age, and backend target.
IT & rolloutThe macOS agent now runs on Intel Macs as well as Apple silicon, from a single universal installer.
IT & rolloutEndpoints move between full and proxy-less capture without reinstalling, including changes made while an endpoint was offline.
IT & rolloutThe local port used for proxy-less collection is configurable per organization.
IT & rolloutAgent logs on endpoints are size-capped so they can no longer grow without bound on disk.
Fixes
macOS network capture no longer conflicts with SASE tunnel traffic.
ChatGPT response capture no longer drops streamed chunks.
Endpoints deployed proxy-less no longer see macOS approval prompts for components they don't use.
Tool calls no longer appear twice in traces.
AI tool inventory now refreshes on its regular schedule instead of only scanning at startup.
Session metrics now respect the selected time window.
Charts no longer shift time buckets across time zones.
Trace detail rendering: non-prompt frames are labeled, embedded scripts no longer break into spurious command chips, and chips render correctly inside tables.
Execution logs no longer appear empty for older sessions.
The full-screen activity view no longer collapses on reload.
Analytics filter values containing backslashes now match correctly.
Python analysis in chat recovers automatically instead of hanging.
Expired dashboard sessions re-authenticate cleanly instead of returning empty chat responses.
Sign-in retries transient identity-provider errors instead of failing.
The Windows tray app no longer opens duplicate instances.
Uninstalling the agent fully removes local credentials and certificates.
June 9, 2026v1.4.4

Managed rollout controls

Not every environment tolerates updates the same way. Some fleets want every endpoint on the newest agent as soon as it ships; others need version changes to follow their own validation and rollout schedule.

Agent auto-update is now a configurable policy, so Origin can adapt to either. Leave it on and endpoints stay current automatically. Switch it off and update timing moves into your own deployment tooling, while endpoints keep capturing and reporting as before. Turning it back on later requires no reinstall.

Improvements
IT & rolloutIT teams can keep selected organizations on their own agent rollout schedule instead of receiving automatic agent updates.
IT & rolloutWindows deployments can omit the full network capture component when supported prompt activity is collected through another approved path.
EndpointsPrompt records keep a model label even when response-side usage data is incomplete.
EndpointsOrigin now reports the AI tools installed on every endpoint, even ones that aren't running full activity capture, so AI tooling is visible across the whole fleet instead of only on monitored machines.
EndpointsEndpoint inventory now finds more AI tools, including the Claude and Codex desktop apps and command-line tools installed in less common locations that earlier versions could miss.
IT & rolloutSupport reports can include attachments up to 250 MB, making it easier to send larger logs and diagnostic bundles.
Fixes
Codex capture no longer records unrelated background beacon traffic.
Fresh Windows wizard installs no longer fail before local agent data is initialized.
Endpoints with no assigned policy no longer restart repeatedly during session processing.
Windows endpoints installed without the full network capture component no longer restart repeatedly when that component is absent.
Perplexity web searches now appear as their own trace turn.
Chat exports no longer loop through forked conversation branches.