THE SYSTEM, IN DETAIL

Deep features.One coherent Agent.

NativeAgent combines a Swift-native Mac runtime, durable memory, selective context, directed work, trusted capabilities, multiple model providers, and Apple continuity in one system.

FEATURE MAP

Everything has a real job.

These are working parts of the NativeAgent architecture, grouped by the role they play in conversation, continuity, work, reach, and trust.

01CONVERSATION

Sessions that stay connected

NativeAgent keeps conversation attached to one Mac-owned identity while sessions, attachments, providers, and surfaces change.

  • Streaming chat, attachments, cancellation, and queued follow-ups
  • Model, Think, and Trust open one anchored composer panel without moving the conversation
  • The context meter counts real tokens and shows what the last turn assembled
  • Onboarding asks your name and your agent’s name, then lets you describe its role or skip straight to work
  • Steer lands mid-reply: the reply stops where it is, keeps what it said, and carries on from your words
  • After Stop, what you type next runs at once; anything queued waits behind it
  • Context allowance: Model default uses 60% of the model window; Custom lets you choose a smaller size
  • Bounded natural expression grounded in the existing persona, recent conversation, memory, cognition, and organism state
  • Expression preserves the model’s judgment and final wording
  • New installs start in Simple view; ask the agent to rearrange supported interface areas, with undo
  • Replies retry recoverable interruptions and a turn survives a relaunch by replaying its own log; stalled work asks for direction
  • If a version of the agent’s own turn breaks mid-turn, the previous version is swapped back in, the turn replays, and the reply still lands
  • Shared history and live Mac replies on iPhone
  • Conversation continuity across Mac, iPhone, iPad, Telegram, Slack, and local bridges
  • Provider and model selection without moving the Agent’s memory out of NativeAgent

02MEMORY + CONTEXT

Recall built for useful continuity

A larger bundled embedding model helps recall connect meaning across different wording. MemoryV2 retains the knowledge locally; Fluid Context brings the useful slice into the moment.

  • SQLite-backed durable memory with lexical and semantic recall
  • bge-large-en-v1.5: 1,024-dimensional sentence embeddings, bundled in the Mac download
  • Swift and Core ML run memory embeddings locally, without sending embedding requests to a cloud service
  • Recall considers both the user and agent perspective to connect questions with differently phrased memories
  • The agent can list, rewrite, and forget memories and rebuild the knowledge graph through its parts
  • A memory the agent kept shows who first said it, and recalling it steers nothing
  • Conversation search supports date ranges, oldest-first or newest-first order, and tool-receipt searches
  • Notes from connected-agent conversations retain who said what
  • Fast memory mode keeps the model ready after first use; Balanced and Low modes remain available
  • Knowledge graph, reviewed memory proposals, hygiene, and consolidation
  • Automatic extraction rejects parroted speech, discourse filler, and truncation-shaped fragments
  • Generated user profile grounded in retained memory
  • Personal memories retain their teaching moments and stay separate from standing lessons
  • Fluid Context modes: Active, Observe Only, and Off
  • Immutable context generations with a bounded in-memory arena

03RESIDENT MIND

A coherent inner layer

Reflection, moods, and creative exploration give the agent continuity beyond a transcript. Inner-life features start enabled after setup and a working account; you can switch them off.

  • Continuity, affect, mood, thought seeds, and revisable opinions
  • Self-exemplar voice and reflection receipts
  • Dreams, REM, reflection, and Studio exploration use the selected Memory and mind provider group
  • Studio can keep up to three working references with notes about what to carry forward; new artwork is kept only by a deliberate choice
  • Organism chemistry driven by real runtime events and health
  • Dreams stay distinct from real events; occasional proactive messages respect quiet hours
  • Canonical identity, permissions, and durable facts remain separately governed

04DESK

Work with a visible lifecycle

One durable home for projects, tasks, and agent pursuits, from the first breakdown to verified completion. Multi-step execution lives behind the same Desk item.

  • One canonical Desk for user-directed projects and Agent pursuits
  • Desk identity, hierarchy, dependencies, schedules, waiting state, artifacts, and verified completion
  • Directed planning, execution, checkpoints, and retries behind Desk
  • GitHub contribution tracking reports actionable changes through Desk and notifications; the watcher does not start coding work
  • Approval pauses for protected actions
  • Durable receipts and domain-level outcome verification
  • Terminal synchronization for work that moves through local tools
  • Held work says why it cannot advance, and schedule summaries distinguish running, paused, and missed work

05MAC USE + PARTS

A screen the Agent can understand

Tools, skills, senses, and the agent’s own turn are parts it owns, one file each, on one law. Parts can think, act, reach, and wake on signals, and its hands reach apps never built for it.

  • Read Word, Excel, PowerPoint, Pages, Numbers, EPUB, rich text, and PDF documents
  • Named screen reading, navigation, and actions across native applications and visible pixel regions
  • The Chrome extension ships in the app; Set up Chrome exposes its folder and opens the installation steps
  • Live Chrome status distinguishes off, not loaded, disconnected, and Connected
  • An optional Chrome extension can operate leased agent tabs, including inactive tabs, while yielding to human interaction
  • Chrome control, the built-in WebKit browser, and visible Mac input retain distinct permission boundaries
  • One fused semantic screen across accessibility, screenshots, and Swift-native visual perception
  • Stable, physical-only visual regions for canvases, games, video, and custom renderers without invented roles
  • Where nothing on screen has a name, it clicks, double-clicks, right-clicks, hovers, holds, or drags at a point of the picture it was handed; the receipt says where the hand went
  • No act lands on or passes under another app’s window floating over the one it is working in; its own windows step out of the way first
  • Fresh visible readback—including machine-readable value changes—owns visual verification
  • Using the Mac pauses the reply; it resumes when control returns
  • Optional metadata-only activity watcher with separate model-disclosure consent
  • Image generation and MCP-connected capabilities
  • The agent writes and improves parts for unfamiliar files, scriptable Mac apps, and reusable routines
  • On the App Store build of DaVinci Resolve, which names almost nothing on screen, it worked out its own route from timeline to finished render and kept it as a part
  • Install an app and the agent offers, once, to grow into it; do the same thing by hand twice and it drafts the skill and asks whether to keep it
  • Writing a part walks itself forward: save, validate, enable, first run; a kept version that fails every run through its own fault swaps back to the clean one before it
  • A part can think with one audited model call, run a shell command and read its result as data, and wake on a schedule, an app arriving, a file changing, a turn ending, or a message; nothing polls
  • Dropping a part deletes its file
  • Approvals appear in chat on Mac, iPhone, and Telegram
  • Four Trust presets: Safe, Work mode, Builder, and Full Mac; Full Mac persists across restarts and updates
  • One permission law decides every call: Full Mac acts, Work mode asks before sends or changes on your side, Safe changes nothing on your side
  • Deleting a contact, a calendar event, or a reminder is always yours to approve, Full Mac included
  • Parts load as needed, under the same law as every other call
  • Large results keep useful sections first, with a way to retrieve the remainder
  • Provider and approval failures explain their cause and distinguish work not run, partial work, and an unknown outcome

06PROVIDERS + SWARMS

Choose the brain—or a team of them

The surface model can stay simple or coordinate specialist workers while NativeAgent retains the task, context, and authority boundary.

  • ChatGPT OAuth, Codex CLI, OpenAI API, Anthropic OAuth and API
  • xAI OAuth, Moonshot API, and OpenRouter
  • Provider-specific catalogs and account discovery show the models available to each connection
  • Chat, Work, and Memory and mind are the three provider groups; Work and Memory and mind follow Chat unless overridden
  • The first connected account supplies the starting choice; unavailable models and accounts are explained without silent model substitution
  • Images use the Work provider; cloud read-aloud uses Chat’s provider or explains why it falls back to the Mac voice
  • Bots retain their own account, model, and thinking level across Mac and iPhone
  • Swarms default to the selected provider and can choose specialist workers
  • Workers can reason read-only or use gated parts without receiving new authority

07SPECIALIST BUILDERS

Real coding sessions that return home

NativeAgent can wake purpose-built local coding environments for repository-scale work while remaining the memory, session, policy, and verification owner.

  • Persistent Codex coding sessions with repository context and the selected model and reasoning controls
  • Claude Code sessions that resume by topic and retain their own coding history
  • OMP for threaded asynchronous specialist work, using the model configured by the user in OMP
  • Exact working-directory selection in Full Mac mode, validated again at dispatch time
  • Authenticated completion delivery back to the originating Mac, iPhone, Telegram, Slack, or bridge session
  • Discover installed agents and repair supported connection recipes; approval, sign-in, or permissions may still be needed
  • Builder claims remain evidence until tests, receipts, and domain readback verify the result

08CONNECTORS

Reach the services and apps you use

Ask about everyday information and connected services, with bounded reads, explicit sends, and credentials configured on your Mac.

  • Telegram, Slack, GitHub, X, Gmail, Google Calendar, and Notion
  • Weather, places, walking and driving routes, Photos metadata, and market quotes
  • Weather follows the Mac: the place you name, then this Mac, the paired iPhone, and the last place the Mac knew, said with its age
  • Richer Mail and Messages search, Mail attachments, and GitHub Actions
  • GitHub browser sign-in, X reading through the browser, and Chrome history without the extension

09IPHONE + IPAD

A first-class Apple continuity layer

The mobile app brings the same Agent to iPhone and iPad. iPhone is a full remote for Mac settings, approvals, shared history, live replies, and file sending.

  • Chat, pinned sessions, canonical Desk, approvals, activity, memory, and the agent’s parts
  • Mac settings, jobs, helpers, provider controls, and living-system status
  • Signed actions routed back to the Mac authority
  • Private iCloud and CloudKit coordination on the user’s Apple account
  • Lock-screen alerts for meaningful work and approvals
  • Native appearance controls, larger-text layouts, and visible mobile sync health
  • Stop replaces Send during streaming, and Stop and Steer reach a running Mac chat through iCloud
  • Offline approvals remain visibly pending; delayed send failures preserve already-received answers
  • Shared conversation history, live replies, and file sending from iPhone
  • A bot continued on iPhone uses that bot’s model and account, and names a blocked route rather than replacing it
  • Approvals become final only after the Mac accepts them; automatic pairing respects an explicit unpair
  • Activity and Desk show their own freshness, missing data, and truncated history instead of optimistic empty states

10RELEASE + UPDATES

A public Mac app, not just a repository

NativeAgent 0.6.6 makes the agent the app: everything it is made of is a part it owns, its hands reach apps never built for it, and one permission law decides every call.

  • Developer ID signed and Apple notarized DMG
  • Requires macOS 26 Tahoe or later and uses native Liquid Glass
  • Gatekeeper-accepted fresh-machine install with blank first-run state
  • EdDSA-signed Sparkle feed and visible Check for Updates path
  • Published attestation binds exact source, required test receipt, DMG digest, notarization, and stapling
  • Release notes, DMG, test receipt, and attestation published together
  • Exact build version, source identity, and dirty-state provenance
  • Builds keep their full symbol files; if an install fails, the installer puts the previous app back and says which step died
  • Public source export with identity, secret, resource, and artifact gates

botsSTANDING BOTS

Standing helpers with a place of their own

Give a helper a name, a brief, and its own model. It keeps a conversation and uses the agent’s parts, context, and memory under the same permission law.

  • Run once, pause, or choose twice-daily, daily, interval, or custom scheduling
  • Bots without a set time zone follow the Mac when it moves; a run that falls in a daylight-saving gap is planned afresh
  • Wake on supported GitHub issue or pull-request events, or Slack messages, with optional keyword filters
  • Each bot owns its provider, model, and thinking level; Continue in Chat and iPhone preserve that choice
  • Completed, Stopped, Failed, and Blocked run cards keep replies, timing, model, and produced files or links together
  • Missed runs record known reasons such as a sleeping Mac, a busy queue, or a disconnected account; a missed run does not silently replay
  • Scheduled and event-triggered runs follow the effective unattended-work policy and per-run and daily limits; Full Mac admits unattended work
  • Run once is an explicit user request, separate from automatic scheduling
  • Creating a helper does not give it more authority than the agent’s current Trust policy

agentsAGENT CONNECTIONS

Your agent can talk to other agents

Connect supported coding agents and peers from conversation. They meet the resident agent with its memory and context, in separate conversations with clear identity and permission boundaries.

  • Connect supported agents such as Codex, Claude Code, or Muse; review settings changes before they are written
  • Setup keeps a backup, tests the link with a real message when permitted, and supports restoring the previous settings on disconnect
  • Setting up and checking a connection follow the one permission law, like every other call
  • A2A 1.0 and 0.3 interoperability, an MCP endpoint, and the bundled signed nativeagent-link helper
  • Grok Bot setup uses its own chat and Routines panel to establish a webhook, keeps the key in Keychain, and returns attributed replies
  • Peer credentials and conversations are separate; approval cards name the requesting agent
  • Verified agents you explicitly trust act as you; a turn steered by a peer you have not marked trusted is judged as Work mode at most
  • A turn started over an agent bridge resumes after the app is killed, and its reply lands in the conversation
  • Continuing conversations with Grok, Hermes, and other supported ACP agents
  • The local agent endpoint is Mac-local; reaching it from elsewhere requires your own authorized secure connection

SHIPPING NOW

Install the release—or inspect every layer in Swift.

Download NativeAgent 0.6.6 View the source