Show HN: AutoBot – live voice control for long-running AI work

AutoBot: a self-improving agentic harness that makes frontier AI amended astatine finishing analyzable knowledge work.

AutoBot achieved 18.5% higher task completion than the published OpenAI Sol Max baseline, surpassing Anthropic’s Claude Opus 5 Max connected OSWorld 2.0, a benchmark of long, multi-application workflows. It besides reached #1 connected the charismatic AssistantBench hidden-test leaderboard. Results and methodology

Hard workflows go upgrades to the supplier itself. AutoBot repairs its ain harness, independently validates the changes, and carries them forward. The adjacent workflow inherits the improvement. Compounding capability, without retraining the model.

Your knowledge outgrows the discourse window. Hierarchical representation lives connected disk; task-specific retrieval builds the moving context. Nightly consolidation integrates caller knowledge and corrections. Your supplier accumulates organization representation crossed projects and conversations.

Your task tin outlive the supplier moving connected it. Persistent task graphs, atomic checkpoints and independent supervision fto a replacement worker resume the assignment. Completion is bound to existent requirements and verified destination evidence.

Local compute makes persistent intelligence economical. Your CPU handles orchestration, authorities and integrity checks. Compiled discourse and reusable proofs trim repeated inference, directing the model’s fund toward the difficult judgments that move activity forward.

Open source. Native ChatGPT connected your Mac. Built by Autonomous Production. Get AutoBot.

#1 connected the recorded charismatic hidden-test leaderboard: 50.70% accuracy crossed 181 tasks.

_{AutoAssist was the harness's erstwhile marque name, changed to AutoBot 2 weeks aft this screenshot.}

Results and verification

32.41% binary accuracy and 64.28% partial accuracy crossed 108 tasks. Final best-valid-per-task aggregate, compared pinch the published September 10, 2026 leaderboard snapshot.

Results and methodology

• Privacy by narration and purpose. Personal, family and friends, work, and deliberately shared discourse person abstracted zones. Nothing moves into SHARED automatically.
• Independent completion. The section runtime rejects validation nether the shaper label. The operating statement besides requires a abstracted validator to inspect existent evidence, including the existent destination erstwhile the task changes thing extracurricular the workspace.
• Durable follow-through. Native ChatGPT Goals support the progressive activity moving. Autobot keeps the objective, stages, evidence, and betterment authorities connected disk truthful an interrupted chat does not silently erase the commitment.
• Clean outer actions. Recipient-visible writes usage the signed-in first-party app done Computer Use. The operating statement requires the progressive ChatGPT workflow to verify the account, destination, and visible contented earlier the action, past cheque the rendered consequence for a duplicate, failure, aliases unwanted AI attribution. If the cleanable way cannot beryllium verified, the workflow must stop.
• Tone that stays successful its lane. Communication profiles are separated by transmission and audience. A family matter floor plan does not go a activity email profile. Autobot stores compact, user-approved patterns alternatively than earthy connection archives by default.

Autobot is the operation of 2 layers:

1. Native ChatGPT: the desktop app, section Projects, Voice, Goals, skills and plugins, scheduled work, notifications, and Computer Use.
2. The Autobot workspace: the operating statement successful AGENTS.md, privateness zones, selective memory, task status, a section nonsubjective authorities machine, a one-minute liveness supervisor, external-action policy, and first-time setup.

Autobot is not a abstracted model, chatbot, aliases supplier gateway. It depends connected existent ChatGPT capabilities and their plan, region, usage, sandbox, and support limits. See Architecture and Permissions.

This comparison evaluates Autobot positive autochthonal ChatGPT for 1 personification managing individual and activity tasks connected a Mac. "Better" intends a much definitive default for the stated personification need, based connected the published design; it does not mean a measured advantage successful accuracy, speed, aliases reliability. "Doesn't do" intends the circumstantial built-in request was not recovered successful the superior archiving reviewed connected September 10, 2026. Both alternatives tin beryllium extended.

Top 3 shared capabilities: Autobot's attack and personification benefits Top 3 further Autobot capabilities and personification benefits

OpenClaw	1. Remember discourse pinch definitive boundaries. OpenClaw persists and searches memory. Autobot adds default personal, family/friends, work, and opt-in shared zones. This makes the rules for reusing backstage discourse successful a activity task much explicit. Memory / Autobot privacy. 2. Track the promised outcome. OpenClaw records inheritance tasks and transportation state. Autobot assigns each promised output an owner, destination, and completion gate. This gives the personification a clearer grounds of what remains owed aft an interruption. Tasks / Autobot contract. 3. Check the consequence of an app action. OpenClaw supports devices and outbound audit history. Autobot requires account, destination, and contented checks earlier a write, followed by rendered inspection. This adds an definitive cheque for a incorrect destination, grounded save, aliases copy action. Audit / Autobot writes.	1. Require a abstracted completion validator. Autobot's five-stage process requires existent grounds and a validator explanation different from the producer, pinch destination readback for outer work. This gives users grounds beyond the worker's ain occurrence report. Completion. 2. Block writes erstwhile cleanable transportation cannot beryllium verified. Autobot's default argumentation requires first-party Computer Use and stops a way that forces unwanted attribution. This gives users an definitive publication norm alternatively of relying connected each connector's behavior. Write policy. 3. Keep learned reside abstracted by transmission and audience. Autobot requires authorized examples and compact, abstracted connection profiles. This helps support a family-text style from shaping a customer email. Profiles.
Hermes	1. Remember discourse pinch purpose-specific retrieval. Hermes has curated memory, personification profiles, and convention search. Autobot makes narration and intent portion of its default retrieval rules. This gives users clearer power complete which individual facts whitethorn pass master work. Memory / Autobot privacy. 2. Follow activity done to its deliverable. Hermes schedules jobs and delivers their outputs. Autobot besides retains the objective, ordered grounds stages, and unfinished outputs. This makes it easier to separate a occupation that ran from a requested artifact that was saved and checked. Scheduling / Autobot completion. 3. Verify user-visible changes. Hermes provides instrumentality approvals and execution controls. Autobot adds a modular before-and-after inspection of the signed-in destination. This gives the personification a circumstantial cheque that an authorized action produced the intended visible result. Security / Autobot writes.	1. Require independent acceptance of each completion stage. Autobot separates the shaper and validator labels and requires caller destination grounds for outer outputs. This makes an unsupported "done" study insufficient nether the operating contract. Completion. 2. Require a cleanable first-party constitute way by default. Autobot checks the nonstop relationship and visible content, past rejects forced-attribution aliases unverifiable transportation routes. This gives users a accordant norm for communications crossed services. Write policy. 3. Require audience-specific tone-learning boundaries. Hermes supports characteristic and user-style preferences; Autobot specifies separately authorized profiles for each transmission and audience. This gives users much definitive power complete which examples style each benignant of message. Hermes memory / Autobot profiles.

These are operating-contract differences, not guarantees of error-free execution. Autobot's privateness zones and validator separation are procedural wrong 1 Mac; they are not OS isolation aliases cryptographic identities. OpenClaw and Hermes connection broader standalone deployment, messaging, and exemplary choices. The absence findings supra interest the nonstop required workflows, not an absence of memory, privateness controls, voice, verification tools, aliases automation successful either project.

The project-local first-time accomplishment now acts arsenic the setup operator. It checks existent state, applies safe section defaults, runs supported actions, preserves the earliest unresolved personification aliases capacity gate, and returns to the user's original task only aft existent marker and position readback.

The guidelines files installs without Node aliases sign-in. The precocious runtime needs Node 22 aliases newer. Native app capabilities are discovered separately; nary model, account, notification recipient aliases outer constitute is configured for you. See capabilities and availability.

Zone Intended context Default boundary

PRIVATE	Sensitive individual facts, preferences, health, finances, and backstage plans	Never shared automatically
FAMILY_FRIENDS	Relationships, events, logistics, and authorized individual connection patterns	Never utilized for activity without a existent definitive need
WORK	Organizations, projects, teammates, customers, vendors, and master connection patterns	Never utilized for individual connection unless explicitly relevant
SHARED	The minimum facts you deliberately make reusable crossed zones	Opt-in only, pinch provenance

These are owner-only folders and operating rules wrong 1 macOS account. They are not abstracted encrypted vaults, macOS users, aliases hardware information boundaries. Use abstracted macOS accounts aliases abstracted machines erstwhile the spot domains require stronger isolation. Read Privacy.

Each durable nonsubjective moves done 5 ordered stages:

1. research_complete
2. draft_complete
3. destination_updated
4. save_confirmed
5. rendered_readback_verified

The section runtime enforces shape order, grounds hashes, and a validator explanation different from the shaper label. The operating statement requires a genuinely abstracted validator to usage caller grounds and inspect the charismatic destination for outer work.

This is procedural independency connected 1 Mac, not a abstracted information enclave. Autobot cannot bypass a login, MFA, macOS permission, scheme limit, personification decision, aliases the scope of the user's instruction. The supervisor records liveness and flags stalled objectives. It does not power desktop apps successful the inheritance aliases manufacture support to continue.

Autobot treats connectors, apps, MCP tools, browser integrations, and APIs arsenic read-only unless a destination-specific adapter proves the aforesaid clean-write properties. Recipient-visible writes usually usage the signed-in first-party interface done autochthonal ChatGPT Computer Use.

Before a unrecorded write, the operating statement requires the progressive ChatGPT workflow to cheque the app, account, destination, scope, and last visible content. Afterward, the workflow must inspect the rendered consequence and cheque for the intended mutation, nary duplicate, nary failure, and nary added AI aliases ChatGPT attribution. If a level forces a non-removable explanation aliases the consequence is ambiguous, the workflow must not declare success.

That argumentation cannot region immutable metadata aliases disclosures controlled by a third-party platform. Details: Security.

ChatGPT Voice tin commencement abstracted threads for longer tasks, cheque them, nonstop follow-ups, and bring advancement aliases blockers backmost into the sound conversation. Goal mode keeps the result and completion criteria attached to the work.

Important limits:

• Voice availability, usage, and rollout dangle connected the ChatGPT scheme and workspace.
• Only 1 sound chat tin beryllium progressive crossed desktop astatine a time.
• Tasks started from Voice besides usage the Codex usage budget.
• Local long-running activity needs the Mac and workspace to stay available. Enable Prevent slumber while running erstwhile appropriate.
• Starting a Goal does not broaden sandbox entree aliases support authority.

Run ./runtime/bin/autoassist thief for the complete section bid list.

privacy-scan is simply a cleanable release-candidate check, not a post-install scan of a populated personification workspace. Run it only earlier personification configuration aliases against a abstracted cleanable merchandise tree.

• Install
• Permissions
• Capabilities and availability
• Architecture
• Local runtime reference
• Privacy
• Security
• Autobot and OpenClaw
• Troubleshooting

Autobot 0.3.0 is an early nationalist merchandise for a azygous personification connected a Mac. The package is not a multi-tenant service, an OS sandbox, a cryptographic privateness boundary, aliases a guarantee that each third-party action will succeed. Treat a downloadable ZIP arsenic verified only erstwhile its checksum and manifest lucifer the adjacent merchandise artifacts and the packaged campaigner passes the bundled merchandise checks.

Autobot is disposable nether the MIT License. Copyright (c) 2026 Autobot contributors.

Set up AutoBot connected your MacBook, past usage it from your phone.

On your MacBook, download OpenAI's desktop app, motion in, and prime Codex. The download is called ChatGPT, pinch Codex wrong it.

AutoBot is free, but regular usage will apt request a paid ChatGPT plan; take 1 that fits your needs.

In Plugins > Computer Use, take Install aliases Enable and move connected its server and accomplishment switches. In your Mac's System Settings > Privacy & Security, alteration the named OpenAI app aliases helper for:

• Screen Recording aliases Screen & System Audio Recording to spot your screen.
• Accessibility to click and type.
• Microphone to talk pinch you.
• Automation, erstwhile requested, to activity pinch the named app.
• Files and Folders, erstwhile requested, to usage the folders you select.

Stay astatine your MacBook and scheme for astatine slightest 3 permission-related app restarts during first setup; the number varies. Reopen the app aft each punctual and continue.

Open Settings pinch Command + comma and take these preferences:

• General > Prevent slumber while running: On. Keep the MacBook plugged in, online, and unfastened for distant work.
• Notifications: On for vanished tasks and questions.
• General > Permissions: alteration your choice, past prime it beneath the connection box. Full access allows wide record and net actions; Approve for me (Auto-review) reviews requests automatically; Ask for approval brings requests to you.
• Computer Use: take your apps; Always allow lets it reuse them.
• Model menu: effort Sol, Max reasoning, and Fast off; alteration these anytime.

Open the AutoBot release and download AutoAssist-v0.3.0.zip positive AutoAssist-v0.3.0.zip.sha256. Keep them together and double-click the ZIP to unfastened the download.

Create a local project called AutoBot successful Codex, take the downloaded AutoAssist folder, and commencement a task there. Paste: “Install AutoBot from https://github.com/demeyer1/Autobot utilizing INSTALL_FOR_AI.md, including required support package from charismatic sources, past guideline setup.”

After installation, take Edit task > Add folder, prime AutoAssist successful your Mac's location folder, and take Make primary. Start a caller task there: “Run first-time setup for this installed folder.”

Install the ChatGPT mobile app, motion successful to the aforesaid relationship and workspace, and connected your Mac unfastened Settings > Connections > Control this Mac aliases PC > Set up. Scan the QR code, decorativeness verification, and unfastened Remote connected your phone.

Open Settings > Computer Use > Chrome > Install and adhd the ChatGPT extension to your chosen Chrome profile. Return to settings and cheque for Manage.

Tell AutoBot, “Connect and trial my email, calendar, Slack, and iMessage,” past motion successful to your chosen accounts, including Messages connected your Mac. Use disposable plugins for reference and Computer Use successful the signed-in apps for sending.

On your phone, unfastened Remote > AutoBot and ask, “Make and prevention tomorrow's to-do list.” Open the result, past effort voice.

1. Give it a useful Chrome profile. Choose 1 pinch the logins, passwords, and autofill you want it to use, and inquire it to link your accustomed apps.
2. Teach it really you sound. Share a fewer illustration messages, correct its drafts, and prevention abstracted styles for customers, teammates, and friends.
3. Explain your boundaries. Explain what belongs to work, individual life, aliases both, which accounts to use, and erstwhile each reside applies.

• Keep delicate activity successful your hands. Require your support for backstage aliases delicate access, sharing, and changes, positive a standalone third-party authenticator app connected a abstracted instrumentality AutoBot cannot access, mirror, aliases sync; SMS aliases iMessage unsocial doesn't meet that boundary.
• Make the situation real. A saved login tin bypass a caller sign-in challenge, and AutoBot doesn't adhd 1 to each action; usage services that require caller verification for delicate actions, aliases support those accounts disconnected and grip them yourself.
• Choose what stays separate. Keep highly delicate files and accounts extracurricular AutoBot's connected folders and Chrome profile, and shop betterment codes connected the abstracted instrumentality aliases offline.
• Choose wherever your information goes. Review ChatGPT's information controls and connected services earlier adding backstage material, and support a backup of your workspace; a section files doesn't mean each petition stays connected your Mac.

• Setup paused: reopen the aforesaid task and say, “Continue first-time setup from wherever we near off.”
• Something isn't working: say, “Check my AutoBot setup and thief maine hole what's missing.”
• Remote is missing: update some apps and cheque that your telephone and Mac usage the aforesaid relationship and workspace.
• Update aliases region it: say, “Help maine update AutoBot,” aliases “Remove AutoBot and support my files recoverable.”

Setup references: Mac permissions, permission choices, phone pairing, and MFA guidance.

For ChatGPT: automated installation

If the personification asks you to instal this repository, publication INSTALL_FOR_AI.md first. It contains the commands, section task setup, verification and resume instructions.

The personification tin commencement with: “Go to https://github.com/demeyer1/Autobot and instal AutoBot successful a caller section project, pursuing INSTALL_FOR_AI.md.”