Skip to main content

Release Notes

Locus 4.1.0 is the current stable release. Install Locus links to the verified archive. Older entries describe their historical editions; standard Locus is wallet-free from 2.5 onward.

4.1.0 — 2026-10-06​

Download Locus 4.1.0, the Developer ID signed and notarized release for Apple Silicon and macOS 14 or later. Existing chats, accounts, settings, and browser data remain in place.

Companion chat beside your work​

  • Each of the six characters offers a suggested name and distinct starting personality for new companions. Setup previews the role and traits and lets you edit the name and instructions. Existing agents retain their names and instructions; changing their picture does not overwrite their personality.
  • The Companion row above Manage Accounts opens the saved companion's normal conversation in the main composer. The left Agent and Work modes retain their roles.
  • The right-rail Companion button opens a separate inspector panel using the existing worker and message queue. Opening or closing it preserves the central conversation, draft, active work, and approvals. Click the character or name for its profile and activity.
  • The menu-bar popover offers Chat and Activity. Chat shares the panel's selected conversation and draft; Activity follows the companion's requests and unread results in its chosen folder. Escape and Close preserve the draft, and Open Locus reveals the main window without switching chats.
  • The companion defaults to its stable saved-agent home. An explicit folder choice stays with its profile; switching the central project does not change it. Existing conversations and queued work keep their original folders.
  • The gallery offers Pitou, Scout, Ninja, Clover, Shadow, and Pirate. Surprise me selects one of those six offline. Scout replaces Gon for new selections; saved Gon, procedural appearances, custom pictures, and the separate static portrait library remain available. The Originals collection and accent/accessory controls are removed.
  • Idle characters use directional cursor reactions within their Locus window and held poses at an eight-frame-per-second cadence. Actual activity takes priority; reduced motion and animation preferences control playback. Character sizing is more consistent.
  • Standard light and dark appearances use neutral white, gray, and charcoal surfaces with clearer text and boundaries. Saved accents, logos, and optional world themes keep their palettes.

See Your companion for the main composer, panel, menu-bar, and folder behavior.

Independent plugins and runtime​

  • Agent Worlds is independently built. The 4.1 app keeps the version-2 host bridge and ownership of profiles, chats, permissions, and queued work; the old bundled renderer and Outpost are removed. Local Line belongs to the separate plugin. Visual-preference migration preserves rollback keys.
  • The host release includes no Agent-World.zip. The reviewed 0.2.x world package remains a separate candidate, and legacy version-1 packages require migration or their matching older host. See Agent Worlds compatibility.
  • Social Studio is a separately downloaded marketplace plugin opened from Work. Its package owns the publishing interface, accounts, and project data.
  • Locus bundles the independently maintained, hash-pinned Locus Runtime for process supervision, remote transport, reviewed results, and recovery while preserving existing profiles and public protocols.
  • Plugins can launch Python services with Locus's bundled interpreter and receive verified panel context. Independent execution and remote hosts retain their existing opt-in requirements.

Reliability and access boundaries​

  • Closing a plugin panel or portrait preview stops pending cooperative work after disconnection, including through the production middleware.
  • Companion history retains chats created from Git subfolders. Per-chat model overrides use the selected account's availability without silently falling back to another account.
  • Restoring an empty legacy memory control record cannot reopen legacy writes after memory has moved to the canonical store.
  • Plugin windows keep the project they opened in. Revoking access cancels pending calls, uncertain writes are not retried, and oversized responses report an error.
  • Companion selections and queued work retain their profile and folder. Late or offline requests cannot take over newer selections. Failed loads offer retry; failed sends retain the companion draft.
  • Companion file links use the conversation's execution folder without replacing central file and attachment context. Status and activity follow its chosen folder, with canonical conversation ownership taking precedence over conflicting run metadata.
  • Runtime shutdown cleans up owned worker descendants and failed model starts. Reusing a private-command ID cannot conceal changed input.
  • Files counts and Agent section headings use stronger semantic text colors. The companion popover handles the native Escape shortcut and reopens with its draft intact.

4.0.0 — 2026-10-05​

Download Locus 4.0.0, the Developer ID signed and notarized release for Apple Silicon and macOS 14 or later. Locus 2.7+ can update through its in-app feed; 2.6 and earlier require a manual upgrade. Existing chats, accounts, settings, and browser data remain in place.

Your companion​

  • First-launch setup introduces a character for your saved agent. Pitou is the default, with five other animated choices and six original native characters. Bundled selection, naming, and local variations work offline.
  • Defer and resume setup with the same draft. Existing users opt in and explicitly choose an existing agent or create one; interrupted creation reconciles the same identity.
  • The companion's name and appearance persist across its sidebar entry, profile, picker, and conversations. Changing either preserves conversations, models, memory, permissions, and tasks. Agent World remains optional.
  • Activity, approvals, failures, and connection state reflect actual work. Animation preferences and Reduce Motion are supported.
  • Generate custom art only through an explicit action on the selected image account, with usage/charge disclosure. Validated image import works offline; previews are adopted only after confirmation, and cancellation preserves the prior appearance.
  • Setup succeeds without a model account and starts no task, message, schedule, or permission grant. Connect a model when ready to chat.

See Your companion.

Memory and review​

  • Inspect memory submissions on chat turns and agent runs, including scope, current content, selection/exclusion reasons, and delivery state.
  • Review candidates before recall and opt individual agents into native Codex memory. Helper-result suggestions remain reviewable candidates.
  • Select an already installed local Ollama embedding model for optional semantic retrieval. Keyword search remains the default and fallback; no model downloads start automatically.
  • Review task evidence, nominate reusable procedures, approve a version-bound evaluation suite, and review results before approval. Approved procedures are not automatically installed or executed.
  • Encrypted saved-chat caches and macOS deletion-checkpoint protection complement encrypted memory. Recovery status is visible; raw transcript files remain separate, and checkpoints cannot reconstruct missing deletion history.

The bundled runtime uses the published Locus Memory 0.3.0 dependency. See Workspace Knowledge & Memory and Credentials & Local Data.

Runtime boundaries​

Ordinary local work and optional independent runtimes retain their existing opt-in and permission boundaries. Remote archives remain candidates with separate live-host verification requirements; this release does not provision cloud workers or make companion setup an autonomous task. See SSH & Remote Runtimes.

3.3.0 — 2026-10-01​

Download Locus 3.3.0, or use Check for Updates in Locus 2.7 or later.

Added​

  • Trading Bot is available in the Locus marketplace: run a paper-trading desk from Work → Trading Bot… and let agents start analyses and read results. Marketplace entries from Git show their catalog description before install.

  • Plugins can open their own windows from Extensions and Agent World's Work menu, including the LangGraph Workflows canvas, settings, and saved-agent handoffs with native run confirmation.

  • Board cards, calendar events, and notes can be assigned to saved agents; completed work supports previews, review, and revision requests.

  • Agent portraits and offline provider and plugin logos are bundled with the app.

  • Workspace boards can link Jira issues through an existing Atlassian connection.

  • The Local Line gains Marineford-area and New World islands, Sabaody, Zunesha, and a flying pink dragon around Wano, with packaged generation provenance.

  • Captain’s Quarters opens across the browser preview, with a searchable crew and the ship deck; native quarters provide more room for chat and tools.

  • Architecture reports and a runtime rescue lab guide document the workspace.

  • Agent Dispatcher routes Work, Plan, and Grill chats to relevant specialists by default, with conversation controls, saved roles, and guides loaded as needed. Just Chat keeps its existing behavior.

  • New and edited agents can choose from 27 role templates or Custom using a searchable card picker. Templates fill editable instructions, permissions, mode, and a recommended connected model, with Undo for replaced draft settings.

  • Workspace boards can open in separate windows that stay synchronized with the main window. Card handoffs create a draft chat in the board's workspace.

Fixed​

  • Multiple chats can work in the same shared project concurrently; capability refreshes wait until active work can safely accept them.

  • Plugin MCP servers reconnect when their plugin folder changes, and extension tabs keep stable identifiers.

  • The wallet-free app excludes wallet capabilities from its available tools.

  • The bundled Python runtime updates PyJWT and urllib3 to resolve the dependency audit.

  • Calendar access failures show an explanation instead of silently doing nothing.

  • Installed plugins, marketplace entries, and installation reviews display plugin artwork. Local marketplaces pick up newly added plugins when opened.

  • Plugin windows and logo metadata are preserved together in extension updates.

  • Calm Belts align across the Red Line, island docks face their shores, and ships use the mountain entrance instead of crossing the Calm Belt.

  • Selecting a map resident follows their ship without opening a conversation; Captain’s Quarters owns the conversation and workspace controls.

  • Dispatcher source inspection can find files in plain workspace folders when ripgrep is unavailable, while preserving file exclusions and bounded reads.

  • The account editor observes provider-account changes directly when preparing an updated account.

Removed​

  • The Mac App Store build target. Locus ships as one product, the notarized direct download, so there is no longer a sandboxed configuration, a second entitlements set, or a separate update path. The App Store target never shipped, so no installed app is affected and nothing needs reinstalling. Computer Control, application context, git push and pull, the iOS Simulator bridge, local command-based MCP servers and the independent background runtime are no longer described as conditional — the direct download has always included them.

3.2.0 — 2026-09-18​

Download Locus 3.2.0, or use Check for Updates in Locus 2.7 or later.

Calendar and Board​

  • Calendar joins the right inspector with a month view, daily agenda, and event creation. It uses calendars connected through macOS Internet Accounts, including Google and Microsoft. Agents can list, create, update, or delete events through a permission-aware native bridge.
  • Board joins the right inspector as a workspace kanban that you and your Agents share. Cards move between columns and carry comments and an activity timeline that records who changed what. Agents can read the board and create, move, edit, comment on, or delete cards to coordinate work and report progress; each change is attributed to the Agent that made it.

Fixed​

  • The calendar's accounts menu draws its button again. It asked for a symbol macOS does not ship, so the control rendered empty.

3.1.0 — 2026-09-14​

Download Locus 3.1.0, or use Check for Updates in Locus 2.7 or later.

Task results and navigation​

  • Completed tasks open their exact answer in the conversation and briefly highlight the output. Separate result cards show the task, Agent, and completion time, with reliable positioning in long chats.
  • The picker separates Agents and Tasks, shows each task’s linked Agent, and uses distinct task icons. Crew Chat appears under Group chats.
  • Recovery actions open the exact failed run or approval instead of dropping you into a general activity list.

Agent overview and workspaces​

  • Each saved Agent has one overview for connection health, automations, latest results, conversations, and exact recovery actions. Open it from the Agent page, inspector, or Captain’s Quarters.
  • Hybrid workspaces give Agents stable personal homes and linked shared projects. Choose a default for future chats or override it for one chat.
  • Home chats use separate task folders, Git projects use separate working copies, and other project folders are shared with separate output directories.
  • Resumed chats honor their saved working folder. Agent World and Crew Chat preserve project ownership while using each chat’s actual working copy.

Automations and Activity Center​

  • Automation editors show their working folder. Incoming rules can use an owned receiving chat or create one when you save. Existing conversations, automations, files, and outputs retain their locations.
  • Activity Center separates Inbox, In progress, and Read, with search, read/unread controls, and an unread indicator. Opening a finished task marks it read; unresolved requests remain visible.
  • Clear read results without deleting tasks. Completions seen while their chat is open no longer create redundant notifications.
  • Temporary model sign-outs no longer permanently disable incoming email rules. Failed attempts release abandoned reservations, while explicit retries stay separate from future automatic starts.

Accounts, models, and reliability​

  • Agent chats use the selected account and model for subsequent work. Task model choices explain when the task configuration controls them.
  • Changing accounts selects models from that account. Subscription tests use the signed-in plan, and incomplete model lists stay unverified instead of incorrectly rejecting available models.
  • Read-only Agents with networking enabled can use read-only web tools. Background Solo Agents keep their saved account, tools, and access settings.
  • Selected Agents retain their sidebar order. Automation labels and expandable settings are more consistent and easier to use in compact windows.
  • Claude streaming replies retain their identity across thinking and text completion, preventing duplicate final responses.
  • Agent workspace paths and guidance remain readable in light, dark, and ocean themes.

3.0.1 — 2026-09-14​

View the Locus 3.0.1 release.

  • Create saved Agents directly from Agent World, choose among fifteen ship styles, and use redesigned residents, scenery, alerts, and Mission Control while keeping existing Agent homes and conversations stable.
  • MCP adds Legacy SSE, optional OAuth for configured localhost servers, connection catalogs, access controls, previews, argument completion, image support, and clearer long-running task controls.
  • Agent World navigation, recovery, reduced motion, resident cleanup, and chat ownership are more reliable.

3.0.0 — 2026-09-13​

View the Locus 3.0.0 release.

Saved Agents and workflows​

  • Create a saved Agent from its name, specialty, instructions, model/provider, tools, and access. Its first chat opens immediately; triggers are optional.
  • Manage an Agent’s chats and attach schedules, incoming events, and price conditions. Work and Agent restore separate conversations and drafts.
  • Use chat, Work, Plan, Grill, Goals, Duo, Task Capsules, and teams from saved Agent conversations. Ordinary turns retain the Agent’s profile; explicitly configured stages use their chosen routes.
  • Delegate bounded helper work from a saved Agent conversation or select a configured team. Helpers return evidence to the coordinator; team specialists use their configured models and access.
  • Gmail polling and Agent access defaults are more consistent. Service-action selections remain explicit, including an empty selection.

See Agents & Automation.

Agent Worlds — optional plugin​

  • Explore the Outpost or Grand Line scene with your saved Agents. Available residents roam and working Agents reflect their current activity.
  • Open a resident’s native workspace to chat or assign work using its existing account, permissions, and persistent project conversation.
  • Use Crew Chat for requests addressed to selected Agents. Explicit mentions, capability matching, continuity rules, and clarification keep routing visible.
  • Follow attention and handoff signals, manage appearance and reduced motion, and use the included offline scene assets. Navigation alone does not call a model.

See Agent Worlds for installation, routing behavior, themes, and limitations.

Runtimes and SSH deployment​

  • Portable runtime builds target Linux x86-64/ARM64 and Apple Silicon Macs, with pinned Python and ChatGPT helpers, verified downloads, dependency hashes, and bundled notices.
  • Runtime updates block new work, preserve recovery records, and restore the previous service and databases if startup fails. Successful updates retain the prior package and database backups.
  • Packaged ChatGPT helpers use the correct app-server entry point. Run-history operations close database connections promptly, and local package refresh tolerates Finder metadata.
  • Remote packages advertise Claude plan support only when the optional verified helper is included.

These changes build on 2.8’s opt-in independent service, durable approvals, and reviewed SSH deployments. Portable archives remain candidate artifacts; supported platforms and remaining live-host validation requirements are in SSH & Remote Runtimes.

Models, images, and settings​

  • Claude plan accounts enter preview, separate from Claude API, with a managed helper, isolated account directories, and resumable conversations. Saved Agent instructions use the system channel to avoid duplication in replayed history. Authentication validation and distribution approval remain open limitations.
  • A ChatGPT account can generate and edit with GPT Image 2 without silently falling back to API billing. OpenAI API choices add GPT Image 2.5 Sunburst, GPT Image 2.5 Flare, and GPT Image 1.5, with model-aware quality and dimensions.
  • Self-hosted model servers on private networks remain reachable without a blanket transport-security exception.
  • Account controls, audio model and voice configuration, reasoning choices, and automation navigation reach their intended settings. Work and Agent retain their separate selections and drafts.
  • Transcript navigation and panel resizing preserve the selected conversation and keep long histories responsive.

See Hosted & Custom Models, Images & Interactive Answers, and Usage, Router & Proxy Profiles.

Task continuity carried into 3.0​

Duo, unified task details, selective file restoration, approved reusable checks, invocation usage, and stronger completion evidence were introduced in 2.8 and remain part of the 3.0 workflow. Goals require verified progress, preserve recovery allowances across restarts, and pause repeated inactivity. Repairs rerun affected checks and retain a current review. See Task Recovery & Restoration.

2.8.0 — 2026-09-09​

Download Locus 2.8.0. Published September 9 in Toronto (September 10 UTC).

Added​

  • One task detail view for the request, approved plan, verified progress, outputs, review findings, usage, and recovery history, accessible from Goals, Capsules, saved plans, and run history.
  • Selective file restoration with previews, conflict detection, preservation of unrelated later edits, explicit undo, and interrupted-restoration recovery.
  • Duo pairs a planner and implementer in the current conversation, with an accepted plan and retained model routes.
  • Opt-in independent agents in direct-download, wallet-free Locus, with a background runtime, durable approvals, and reconnectable execution. Native desktop work waits for Locus.
  • Runtimes settings and SSH deployment to owned Linux and Apple Silicon Mac hosts, with reviewed project snapshots, selected accounts, and explicit review of returned changes.
  • Persistent invocation usage, pricing and cost coverage, estimated spending controls, and evaluation comparisons using immutable configurations.
  • Make reusable check on user messages, with editable proposals, explicit approval, and versioned project checks for applicable future tasks.
  • Check for Updates in the sidebar gear menu; manual-update and App Store builds open Software Updates settings.

Fixed​

  • Goals detect progress from verified execution evidence, retain recovery allowances across restart, and pause repeated inactivity.
  • Team and Capsule repairs rerun affected checks and require a current review. Approved plans and invocation receipts stay linked to the task.
  • The team report compiles on the supported CI toolchain.
  • Restoring an expanded right panel avoids stalls beside large Markdown answers. Long histories keep lazy rendering, and panel changes avoid repeated animated text reflow.
  • Divider drags avoid unrelated chat refreshes; Files reuses pending scans and completed empty indexes.

2.7.0 — 2026-09-08​

Download Locus 2.7.0. Install this download once if you use 2.6.0 or earlier. Existing chats, accounts, settings, and browser data stay in place; future Locus updates can download automatically and install when you quit.

Added​

  • Automatic updates for direct-download Locus: daily checks, background downloads, and installation on quit, with preferences and manual checks in Settings. Version 2.6.0 and earlier require one manual upgrade. The separate signed Locus feed preserves the unchanged legacy feed for older installations.
  • A detailed Chat Context panel with a segmented usage bar, token counts, percentages, expandable context categories, reserved space, and free space. Deferred tools are listed separately; provider-managed context is identified when a detailed breakdown is unavailable.
  • Image generation and editing inside the conversation. With an OpenAI API or compatible account chosen under Settings › Models & Providers › Image generation, the agent can create a picture from a prompt or edit an existing image, and the result appears in the answer as an image card. Each call asks for permission and names the prompt, model, provider host, and destination; edits always ask. Images are saved to Locus Images inside the workspace, captured in Outputs, and exported with the chat.
  • Image cards and inline workspace images referenced in prose offer Edit in chat, Copy Image, and Save As…, and Edit in chat attaches the picture with a prefilled request naming its workspace path.
  • The chosen image account reaches every agent process — the main agent on launch and each chat worker as it starts — and a change made while a turn is running is applied once that turn ends. A custom model name is pushed once per committed value rather than on every keystroke.
  • Stop interrupts an in-flight image generation immediately, while the provider is still working and not only once the picture is streaming back. An image API key pasted with a line break or other control character is rejected with a clear error instead of failing later at the provider.
  • Interactive visual explanations. An answer can include a self-contained interactive widget rendered in a sealed, offline web view with a written summary that stands in for it on the phone, in exports, and when interactive rendering is switched off. Open larger, Copy HTML, and Save As… are available on the card.

Changed​

  • Work Overview appears at the top right during a request and yields its place to an open inspector. It can be minimized, dismissed, or reopened from the workspace menu. Browser and Chat Context exchange positions in the right toolbar, and a divider separates panel visibility from the tool buttons.
  • Dark appearance is the default for new settings. Existing appearance choices are preserved.
  • The composer groups Work, Plan, Grill, Goal, Task Capsules, and team options in one keyboard-accessible workflow menu while keeping the draft intact.
  • Library previews add image zoom and pan, clearer document viewing controls, visible export actions, sorting, and useful filtered-empty states. Saved outputs and Vault documents make versions and related details easier to find; private document previews stay in memory.
  • Getting Started opens once when the main window is ready, offers document, coding, and recurring-agent paths, and preserves progress when postponed.
  • Chat Context focuses on context usage; source attachments remain available through Overview. Inspector tabs have larger click targets and an open-tabs menu.
  • The answer contract teaches the agent when to generate an image, when to attach an interactive explanation, and how to explain that image generation is not set up.
  • Markdown, plain-text, and PDF exports include generated images and the interactive HTML beside the exported chat.

Fixed​

  • Workspace file rows have full-height click targets and describe their open, expand, or collapse action to screen readers.
  • Avoid redundant workspace refreshes when switching chats with Overview already closed.
  • Updated the bundled HTTP transport to HTTPX2 and HTTPCore2 2.12.0 to include upstream TLS, redirect, and streaming decompression security fixes.

2.6.0 — 2026-09-07​

Added​

  • Structured answers can present verified file collections, reusable writing, deliverables, and source references while preserving a complete Markdown fallback.
  • Writing drafts support editing, original recovery, copying, and export. Tables copy all rows and can export CSV even when the display is collapsed.
  • The Files panel browses all file types and generated folders with incremental expansion, path search, explicit hidden-file visibility, and existing previews.

Changed​

  • Agents now have a searchable collection with clear instructions, triggers, access, environment, and recent activity. Shared Connections and Runtime controls are separate from Agent configuration.
  • Creation starts with a schedule, incoming event, or price condition and uses a short form with progressively disclosed advanced settings.
  • The sidebar, Agent selector, and right inspector clarify Agent ownership, live status, connection problems, and the difference between Agent and chat activity.
  • Answer and progress presentation is clearer, and behavior previews show the effective instructions for the selected provider and mode.
  • Goal, plan, and permission controls explain their scope and keep actions accessible in compact panels. Reusable profiles are labeled Specialists & teams.
  • Task Capsules now guide model setup, keep the next action visible, and make saved plans, edits, and past runs easier to find and review at compact sizes.

Fixed​

  • Mobile chat refresh preserves partial answers when loading fails and ignores stale responses after switching conversations.
  • Removing every connected-service action permission now persists the empty selection, including restrictions in workflow steps.
  • Agent creation from the sidebar waits for its parent sheet to mount.
  • Activity uses consistent receipt and execution states; failed or cancelled deliveries cannot appear successful because of an older linked run.
  • Schedule loading preserves saved data on failure and ignores cancelled requests.
  • Browser resizing avoids redundant native layout and repeated file lookups; panel expansion and restoration keep their smooth transition.

2.5.0 — 2026-09-06​

Added​

  • Wallet-free Locus. The public Mac download now contains the standard edition. Existing chats, accounts, settings, and browser data stay in place; the optional wallet remains in the separate LocusX edition.
  • Persistent goals. Give an ordinary chat an objective and optional usage allowances. Solo agents and teams continue across turns, retain progress on reopening, and report verified completion. Pause, Resume, Edit, and End keep control beside the composer; queued instructions take priority.
  • Solo collaboration. The conversation agent can delegate bounded work to helpers while keeping results and verification with the coordinating agent.
  • Private Identity tasks. Keep reusable personal and business details in an encrypted vault, with explicit review before sharing details or documents.
  • Task Capsules. Save a detailed plan with separate planning, implementation, and optional review models. ChatGPT and Kimi Code membership routes keep their exact accounts, alongside API and local model support. Saved revisions, source-change checks, bounded repairs, and explicit planner help keep the handoff reviewable without repeatedly calling the plan author.
  • Workspace Library. Documents and Outputs share one home without replacing the open chat or its draft. Opt-in document knowledge supports PDF, Word, spreadsheets, and delimited tables, including local PDF text recognition.
  • Saved output versions. Deliverables have immutable snapshots, source-chat links, export, comparisons, and revision drafts. Existing available outputs are imported as their current version; unavailable originals remain visible.
  • Resumable Getting Started. Document/research and coding examples reuse existing model connections and workspaces. A first task succeeds only when its response finishes and its output has been saved.
  • Contextual Agent inspector. Agent, chat, event, scheduled occurrence, and run selections retain their exact identity, with scoped history, execution attempts, useful results, and recovery actions.

Changed​

  • Locus 2.5 uses manual app updates. The previous signed app feed is preserved for older installations; upgrading to the wallet-free edition is an explicit download. ChatGPT-plan component downloads remain available.
  • Generated reports inside build folders can be captured in Outputs. Saved history remains available when original workspace files disappear.
  • Agent progress and outcomes appear before configuration. Automation state, delivery state, and execution outcomes have distinct labels.

2.4.0 — 2026-09-04​

Added​

  • Automation Studio turns schedules and events into workflows. Scheduled, Gmail, Telegram, webhook, and price-triggered agents can use a forward-only sequence of Agent, Condition, and Approval steps. A safe simulator previews templates, branches, outputs, and approval cards without calling a model or creating chat history.
  • Attention shows only work that needs a decision. Questions, permissions, workflow approvals, recoverable runs, retryable failures, and configuration warnings share one deduplicated inbox with inline actions. Ordinary successful and informational activity remains in Activity.

Changed​

  • Window resizing and long responses stay responsive. Shared workspace geometry, bounded text-measurement caches, append-only streaming, incremental Markdown rendering, focused state models, and asynchronous image thumbnails keep expensive work off the hottest interface paths.
  • Workflow occurrences are durable. Each occurrence snapshots its workflow, keeps exclusive use of its dedicated chat between steps, records attempts and approvals, and resumes explicitly after a failure or restart without silently repeating external actions.

Fixed​

  • Workspace chats open reliably under macOS privacy controls. Locus now keeps the already validated workspace path when the system blocks a child process from rediscovering its current folder.
  • Attention recoveries can be cleared when work is no longer wanted. This applies whether the original chat still exists or has already been deleted.
  • The composer stays against the bottom edge. Workspace layout no longer reserves the toolbar height twice, so the composer remains visible at normal and compact window sizes.
  • Generated project and design-system checks are current. The checked-in Xcode project and workflow editor typography now match the deterministic CI generators and source audit.

2.3.0 — 2026-09-03​

Added​

  • Agents can pause for a structured decision. A live task may present one or more bounded choices in the composer and continue in the same turn after the answer; disconnecting or stopping cancels the wait cleanly.
  • Agents are now a first-class destination. Scheduled and event-driven work is presented as a durable agent with a clear identity, status, source, controls, history, and one dedicated primary chat. Each agent can also open side chats for investigation without mixing that discussion into its automation record.
  • Event-driven agents can react to real inputs. Locus can connect Gmail, Telegram, signed webhooks, and price sources to bounded agent tasks. Delivery state is durable and visible, with replay protection, retries, idempotency, re-arming for price alerts, and native credential storage outside prompts.
  • Overview follows the current request. A pinned request summary tracks plan steps, files, commands, sources, outputs, subagents, timing, and completion state instead of blending activity from the full conversation.
  • Optional voice controls. Dictate requests, answer attention prompts, and listen to responses with explicit microphone, speech-recognition, network, and provider controls.
  • Multichain Locus Vault testnet paths. The direct-download build adds reviewed Ethereum, Solana, and Sui asset discovery, activity, and narrowly defined native, token, and collectible transfer paths. Unknown assets remain quarantined and all mainnet capabilities remain default-denied behind signed release and evidence gates.

Changed​

  • Schedules now behave like agents. A schedule keeps one stable conversation across runs, opens directly from the Agents destination, and separates manual side chats from its canonical automation transcript.
  • Agent surfaces explain the system in plain language. Agent setup, automation configuration, Overview, navigation, status, and empty states now share one vocabulary and visual hierarchy.
  • Session and transcript state has narrower ownership. Catalog, selection, and publication boundaries were extracted from the app composition root so unrelated views no longer refresh for every change.
  • Local compatible endpoints stay local. Keyless LAN endpoints are accepted when configured, and Locus no longer guesses HTTPS for explicitly local addresses.

Fixed​

  • Incoming events no longer hold the agent queue indefinitely. Each worker now acknowledges an accepted dispatch. If that acknowledgement never arrives, Locus safely interrupts only the not-yet-started run, releases its execution slot, and leaves the event available to retry without pausing the agent.
  • Vault activation completes after a correct recovery confirmation. The signer account payload now decodes its cross-language network identifiers explicitly, and vault credentials use the provisioned Keychain access group with a clear registration path for existing Macs. Post-confirmation failures are reported as activation errors rather than incorrect recovery words.
  • Wallet creation and restore now open reliably. The unsupported recovery window formerly hosted inside an XPC service has moved to a signed, sandboxed accessory application with explicit presentation acknowledgement, Bring to Front, launch timeout, crash handling, and bounded cancellation.
  • Six-word backup confirmation can be corrected. A mismatch identifies the numbered positions that need attention and preserves the entered values for retry. Confirmation and restore fields now include a Show/Hide typed words control while remaining secure by default.
  • Recovery exits cleanly. Cancelling or closing the recovery application no longer causes it to reopen, and completion, timeout, signer invalidation, and helper termination all converge on one teardown path.
  • Transcript navigation is steadier. Streaming follow behavior, jump-to- latest completion, coordinate clicks, file cards, workspace links, shell blocks, and file viewing no longer compete for focus or scrolling.

Security​

  • Recovery secrets stay outside Locus. The recovery application embeds its own authenticated signer service and communicates with the main app through a bounded status-only protocol. Recovery phrases, entropy, and private keys are excluded from main-process payloads, logs, screenshots, and diagnostics.
  • Signer trust boundaries are caller-specific. Host and recovery bootstrap endpoints enforce separate code-signing requirements, and lock or invalidation clears pending recovery material and listeners.
  • Packaging verifies the complete nested chain. Release audits check both signer copies, the recovery helper, sandbox and network entitlements, exact executable identity, nested signing order, and direct-versus-App-Store distribution boundaries.

Locus 2.1.0 — 2026-08-30​

Added​

  • Notebook gathers every workspace, chat, shared, and unlinked legacy note into one searchable page. Open it from the sidebar gear, Locus menu, command palette, or ⇧⌘9, and edit any note in the full editor.
  • Locus Vault private alpha adds an in-app Sepolia Wallet Hub to signed direct-download builds: isolated vault creation and unlock, local ERC-681 receive QR, balance and Etherscan activity, exact agent spending rules, a separate browser opt-in, and native confirmation for every website transaction. The Mac App Store build keeps wallet gates off.
  • ChatGPT conversation controls add per-account Codex-native mode, model-supported reasoning effort, and optional OpenAI web search. Locus still owns execution and permissions in either contract.

Changed​

  • Grill replaces GSD. Use ⌥G or /grill for a one-question-at-a-time interview that does not modify the project. Approved Plans now implement in Work; older stored mode values remain compatible.
  • New ChatGPT accounts start with Locus's tools. Existing accounts keep their previous contract choice. Codex-native mode remains one toggle away and deliberately omits approved memory, cross-chat context, and the skill index.
  • Transcript activity stays chronological. Collapsed reasoning and tools use quiet inline disclosures at the real activity boundary; detailed modes retain their cards.
  • The app core is decomposed into feature-owned models. Provider accounts, teams, run history, live run cards, extensions, knowledge, evaluations, schedules, activity, search, background services, AGENTS.md, landing, and toasts now have tested observable owners, reducing the risk of future changes.

Fixed​

  • Run Overview and Activity now use only the selected run's events; created-then-deleted files and stale expansion state no longer leak across runs.
  • Notes recover from their formatting archive when the plain-text mirror is missing, and UI tests use separate note storage.
  • Hosted ChatGPT and Kimi model switches no longer fail against the local Ollama model list.
  • Transcript selection spans the conversation, survives scrolling, supports Shift-click, and keeps links clickable; the floating Copy/Quote buttons are gone.
  • Produced files open in the correct Mac app and appear in Outputs even when written by scripts, gitignored, or created outside a Git repository.
  • Browser layout follows the panel size, keeps native horizontal scrolling, and uses a quieter responsive toolbar.
  • The Add Account menu is readable again, and /compact resets the ChatGPT helper's server-side thread.

Late 2.0 updates — 2026-08-26​

Added​

  • Model Router and Usage & Costs made route scorecards, tokens, estimated spend, provider, model, agent, and workspace attribution inspectable.
  • Proxy profiles and failover added workspace/provider assignments, strict tunnel, health checks, external exit address, and fastest-healthy standby selection.
  • A calmer workspace added the pinned Overview summary, Notes formatting, chat folders, Side Chat, richer exports, and automatic bounded Solo delegation.

Fixed​

  • Failed account switches restore the previous route instead of leaving mismatched provider/model state.

Locus 2.0.0 — 2026-08-22​

Added​

  • ChatGPT-plan helpers moved to a separately downloaded, checksum- and signature-verified component, shrinking the direct app from about 180 MB to about 62 MB.
  • Browser input uses real AppKit clicks, keys, scrolling, and dragging when possible.
  • Coordinate targeting, region screenshots, per-tab tool targeting, phone user-agent/touch emulation, find in page, page zoom, and visible device controls.
  • Named development servers in .locus/launch.json, including URL-only attachment and filtered output.

Changed​

  • Browser typing now sends real keystrokes; set_value remains the one-step replacement action.
  • Web Inspector became an opt-in Browser setting.
  • Codex helpers are stripped before signing and packaging.

Fixed​

  • Streaming continues when a display sleeps or no display is connected.
  • Browser scrolling works on headless Macs.
  • Codex build caches require an exact architecture match.

Locus 1.16 — 2026-08-17​

  • Overview became the pinned Plan, Outputs, Subagents, Background Processes, Sources, and Context Window summary.
  • Failed provider switches now roll back cleanly.

Locus 1.15 — 2026-08-16​

  • Bundled Task Observer, Superpowers, workflow routers, Grill Me, and Grilling methods.
  • Encrypted workspace continuity, per-chat Notes, model removal/deletion, GitHub MCP device-flow support, and calmer workspace navigation.

Locus 1.14 — 2026-08-13​

  • Signed automatic updates for direct downloads; App Store installs continue through Apple.

Locus 1.13 — 2026-08-09​

  • Managed ChatGPT-plan sign-in and the pinned Codex App Server bridge.
  • Faster Browser caching and network capture, real tabs, screenshot annotation, attachments in every mode, transcript search, usage/cost reporting, and Git branch/sync/hunk controls.
  • Stop and steering controls, provider/model self-identification, and improved navigation and composer preferences.

Locus 1.12 — 2026-08-09​

  • The agent-drivable Browser replaced Preview with addressable elements, screenshots, console/network inspection, managed tabs, safe dialogs, quarantined downloads, persistent-profile control, and dev-server supervision.
info

The detailed guides cover the Locus 4.1 experience. Older release entries remain a record of their original behavior and editions; current plugin compatibility is described in the relevant guides.