hunch / intentional agent / hunchagent.io

hunchintentional agent

it knows you. it understands the moment. it acts on your behalf.

In the current TestFlight beta, talk and write to your companion and send requests to your connected computer. Classic chat remains optional. You choose which memories Hunch may use.

categoryintentional agent
principlequiet authority
lineageowner credential
The Hunch moment

A real signal becomes a meaningful action — within your rules.

This is not mind-reading magic. Hunch gains its edge from real signals, learned connections and a long-term understanding of you. One example: a new file shows up that belongs to a decision from three weeks ago.

New file detected: “Launch-Claims_v3.pdf”observer

signal · local sensor

Belongs to intention “Launch” — decision from 31 julobserver

memory · project

Intent: check claims against product truth · confidence 0.91observer

intent

Thread “Launch check” started · 3 deviations foundexecutive

thread · executive

Proposal: set two claims to “experimental” — approve?you

authority gate

Approved · PRODUCT_TRUTH.md updated · way back availableexecutive

result · audit
1 thread · 1 approval · 0 unaskedlearned: check claims before publishing
Not another AI agent

How it differs from ordinary prompt agents.

prompt assistant

ordinary assistant

no seal · no provenance
identitysession id
memorythis chat
scope—
authorityon/off
owner credential

hunch

hunch
identityHID · XXXX XXXX · secure enclave
memorybrain · local · scopes
scopeglobal · project · thread
authoritygovernor · gate · audit · undo

Ordinary assistants

  • wait for a prompt
  • start from scratch in every new chat
  • mostly know the current conversation
  • need to be told again and again how to work
  • process tasks one after another
  • tie identity and features to one vendor
  • treat autonomy like an on/off switch

Hunch

  • also reacts to relevant changes, not only to prompts
  • has a lasting identity and relationship with you
  • connects memory, goals, patterns, working style and context
  • remembers how you want to be addressed and supported
  • pursues several threads in parallel
  • is separate from the interchangeable model
  • distinguishes perceiving, proposing and acting — and is explicitly allowed to do nothing

A good intentional agent is not as active as possible. It also recognises when no intervention makes sense.

Personal model & anticipation

What makes Hunch anticipate.

The personal model connects only activated, controllable sources — among them memories, conversations, tasks, goals and intentions, decisions, people and topics, working rhythms, apps, files and projects in use, recurring routines, open loops and the results of earlier actions. “Hunch reads everything” is explicitly wrong — browser, clipboard, file, app and screen context are optional, locally controllable sensors.

0006121824
R

Routine — what usually runs at this hour

today
F

Follow-on — where focus leads next

today
C

Recurring — what is due again

today
M

Momentum — what is gaining weight

today

Routine Today

What usually happens at this time of day? An hourly profile of your topics.

Follow-on Today

Where does the current focus usually lead next?

Recurring Today

Which recurring topic is due again right now?

Momentum Today

Which topic is gaining so much weight that it will probably continue?

Hunch also detects topic spikes, dormant projects resurfacing, unusual activity times, new bridges between previously separate topics and relevant old decisions in the current context. These four kinds of anticipation are pure statistics over your own memory — no extra model call, no cost. The “pattern of life” runs in a single runtime: hour profile, transitions, cadences and the knowledge graph live under /muster and /graph.Today

Brain, intentions & open loops

A living personal memory — not an empty dashboard.

Brain holds memories, conversations, tasks, goals, working preferences, relevant people and topics, momentum, open loops, connections as well as results and failures. By default it lives locally on your device.

128memories
42conversations
7open tasks
3intentions

Prefers short answers, direct, no follow-up question after every reply.

from you · global · 12 aug

Project “Launch”: check claims against product truth before publishing.

derived · project · 19 aug

Thread “Berlin”: Authority Gate before purchase, budget € 60.

from you · thread · today
Task ≠ intention.A task is a single step. An intention is something you want to achieve or change. Hunch does not decide your goal.
You set purpose and goal — Hunch decides the path within its authority. Multi-day intentions run: a plan of steps, steps as threads, across days and restarts, failure → replan or pause and ask.Today
Hunch gets better with you — under control

Self-improvement is not a vague “the AI keeps getting smarter”.

run 01 · export · check · mark · ok
run 02 · export · check · mark · ok
run 03 · export · check · mark · ok
condensed · 3 successful runs
skill · visibleclaims-check

Exports the claims, checks them against product truth, marks deviations.

editableremovableauthority required

Understanding Today

Recurring experiences are condensed into reliable knowledge. Fleeting content and suspicious instructions are filtered.

Assumptions Today

When an intention changes, Hunch can correct or withdraw assumptions it derived itself. Memories you wrote yourself stay untouched.

Its routinesToday

If a complex tool sequence succeeds repeatedly, Hunch can turn it into a visible, editable skill — available again next time.

Hunch doesn't simply remember more. It learns which context really counts, which paths work and when an earlier assumption no longer holds.
No autonomous training of its own foundation model, no uncontrolled self-programming, no rewriting of the entire source code. Successful routines become controllable skills — visible, editable and removable. Pattern of life and knowledge graph run in one runtime (loop “graph”, incremental, no model call).Today
Personal working style & memory

Hunch learns not only who you are — but how it should work with you.

It remembers not only facts but the way you want to collaborate: whether you want short or detailed answers, whether Hunch should act directly or explain first, when questions make sense, which decisions Hunch may take itself and which actions must never happen automatically.

GLOBAL · TODAYPROJECT · TARGETTHREAD · TODAYSESSIONhunch
01

global — identity · goals · style · boundariestoday

02

project — knowledge · rules · roles · statetoday

03

thread — task · progress · approvalstoday

04

session — dialogue · short-lived · assumptionstoday

Global levelToday

Identity, goals, communication style, basic working preferences, permanent boundaries.

IdentityGoalsStyleBoundaries

Project levelToday

Project-specific knowledge and rules, roles, files, decisions, current state.

KnowledgeRulesRolesState

Thread levelToday

Concrete task, working context, progress, tools, approvals, result.

TaskProgressApprovalsResult

Session contextToday

Current dialogue, short-lived information, assumptions not yet confirmed.

Dialogueshort-livedAssumptions
Each memory shows, where possible, source, scope and time — whether it came from you or from Hunch, whether it is global, project- or thread-bound, and whether it can be edited, forgotten or deleted. Explicit instructions from you take precedence over derived preferences.
Example rule: “Be direct. Carry out safe, reversible tasks without unnecessary questions. Push back only on a real error, risk or irreversible consequence. Then report the result.” — The Authority Gate always stays active: a “don't ask” preference can never override critical approvals, owner rights, irreversible actions or security boundaries. Memories carry their level (global, intention, thread) and can be moved — in runtime, app, MCP and CLI.Today
Observer & Executive

One watches quietly. One acts — only within your rules.

Hunch doesn't think out loud all the time. A quiet Observer recognises what might become relevant. Only then does the Executive take over.

Observeralways on · small · cheap
signalbundlememory & patternsintent + confidencemay: do nothing
threshold 0.85 · wakes the executive
Executivewoken · large · safeguarded
plantools / threadauthority gateactverifyresult
Always on · small · cheap

Observer

Watches only activated events, never answers. Bundles signals, checks memory, patterns and working preferences, estimates intent and confidence — and can decide to do nothing.

Woken on demand · large · safeguarded

Executive

Wakes only at sufficient relevance, plans the next step, uses tools or starts threads, checks authority, explains its why, acts and verifies the result.

Observer and Executive may use different providers and models independently of each other.Today

History + presence

One shared history, targeted delivery.

Console, Telegram and web conversations mirror into the same history. The app archives its chats in the Brain, shows conversations from other surfaces and can bring their last 30 lines in to continue. Presence routes nudges and approvals to a reachable surface.

  • App / Brainarchive · today
  • Runtimehistory · today
  • Terminalcli · today
hunchowner credential
HID · XXXX XXXX
  • Telegramhistory · today
  • Webhistory · today
  • History + presenceshared · today
Shared memory and one shared history exist for the app, runtime, CLI, web and Telegram.
Presence routes nudges and approvals to the active surface first, then to open surfaces or push.Today
Threads

You keep thinking, Hunch keeps working.

Threads are parallel, persistent lines of work within your relationship with Hunch. While a longer task runs, the main conversation stays free immediately — an ordinary assistant would block the dialogue. Every thread has a clear goal, its own context, its own tool rounds, approvals and a result that returns to the original channel.

shared historyConversations are mirrored; each thread keeps its own working context.
Prepare launch

Next step: check the last claims against product truth

Berlin

Next step: compare connections, Authority Gate before purchase

Research

Next step: bundle sources and mark assumptions

Running

The thread works on its task and shows progress and runtime.

Waiting

The thread is parked on an approval and continues afterwards — or finds another way.

Done

The result is ready and returns to the surface the task came from.

States: running, waiting, done, failed, cancelled — persistent across runtime restarts. Threads run in the runtime; app, CLI, web and Telegram mirror conversations into one shared history, while presence routes delivery Today.

Authority Gate

Hunch may recognise intentions. It does not invent authority.

The Authority Gate consists of several layers of protection you can inspect one by one. Hunch proposes; anything with an effect is yours to approve.

Elevated RiskAuthority Gate

Hunch wants to submit the booking for real.

This action creates external costs and cannot be undone automatically.

Owner CredentialAlex RothHID 1683 8762 8792 7165
Secure Enclave · verified
Scope
Thread “Berlin”
Provider
ÖBB
Amount
€ 49.90
Log
active

Specimen · anonymised data · receipt follows the decision
Identityowner binding
Governorrisk and boundary
Second opinionconcrete action
Preview & consentvisible consequence
Execution & auditrecorded outcome
Way backwhen technically proven

IdentityToday

The owner device creates a private key in the Secure Enclave — biometrically protected, not exportable, not in backups. Your HID is derived from the public key.

Authority & riskToday

The Governor distinguishes risk classes, budgets, limits, always-ask rules and circuit breakers.

Second opinionToday

A separate review model may only clear specific, local, reversible Hunch write steps. Critical Hunch tools are never approved automatically. Codex is separate: read-only is the default; write access in the chosen working folder must be selected deliberately and then runs without per-command approval. Hunch does not offer full access for Codex.

Preview Today

Before protected actions you see tool, arguments, target, risk and reasoning. You approve or decline.

Execution & auditToday

Action, decision, result and cost remain traceable in the append-only log.

IsolationToday

Guests get a separate context — neither your personal memory nor your tools.

Impulse brake Today

Minimum spacing plus a daily limit for impulses; three failures of one tool lead to a pause.

Undo & boundaryToday

Remembered facts and created tasks can be taken back; the boundary — what Hunch never does without asking — is visible in one place. Every action sits in the logbook with its reasoning; the way back covers everything recoverable, shell and agent runs are honestly marked as not reversible; every device has its action boundary.Today

Hunch ID Beta

You confirm an approval directly in Hunch with Face ID or Touch ID. Hunch ID uses the device key in the Secure Enclave. Declining remains possible without biometric confirmation. This confirms the device owner; it is not an independent proof of a globally unique person. Full physical-device acceptance testing remains open.

Hunch may perceive, formulate an intent and propose something — but act only with matching authority. Deleting, buying, sending externally and publishing are specially protected.
Before an owner identity is set up, a less strict mode can exist for compatibility. The cryptographic lock is not automatically active before setup is complete.
Models, skills, MCPs & harnesses

The model is interchangeable. Hunch stays.

Bring your model, keep your hunch. Hunch's identity, memory, authority, working preferences and skills are not tied to a model provider — three independent layers.

observer · socketClaude Haiku 4.5small · always on · 0.85 threshold
executive · socketClaude Fable 5large · woken · authority required
claudeopenaigeminikimideepseekqwenmistralgrokgroqapple on-deviceollamalm studiollama.cppopenai-compatible

Hunch Today

Identity, Brain, intentions, working style and authority. The lasting core.

ModelsToday

Interchangeable engines for Observer and Executive — selectable independently.

claude · openai · gemini · kimi · deepseek · qwen · mistral · grok · groq · apple on-device · ollama · lm studio · llama.cpp · your own OpenAI-compatible endpoints.

CapabilitiesToday

Provider-neutral via SKILL.md, Python skills, MCP servers, function schemas, REST connectors, built-in tools and external agent harnesses (Claude Code, Codex, Gemini CLI, Aider, your own CLIs).

Skills and tools are not tied to the model provider. Hunch unifies open formats and adapters — MCP, function schemas, Markdown and Python skills, REST and external agent harnesses.
Any provider can be connected if the model or tool is reachable via a supported standard or adapter. A proprietary plugin package may need an adapter. Additional providers (OpenRouter, Together, Fireworks, Cerebras, Perplexity, Bedrock) — registered in runtime and app, OpenAI-compatible with your own key.Today
One Hunch, several entry points

Not just an iPhone app.

The runtime connects Hunch to your computer and permitted tools. In the current iPhone beta, the companion is the main view; enable classic chat under Settings → View. Both use the same conversation, without a new setup. Terminal, Telegram, WhatsApp, MCP, HTTP and WebSocket are further entry points — not every surface has the same feature parity.

Delivery status · 1 October 2026: iOS 1.5 is the confirmed public release; Apple readback at 22:28:19 CEST. TestFlight 1.6 (234) is available for external testing; Apple readback at 22:24:19 CEST. Companion as the main view, chat optional. Additional Gemini spoken command is experimental on iOS 26+; actual Voice and background device tests remain open. Nine animals in the catalogue, three experimental bird models available; the other six are not finished. Meta live vision and the new Health runtime are not part of build 234. No new direct iPhone installation confirmed. Mac and Windows still have no publicly verified download; Mac 231 is only in the private development update channel.

TestFlight betaiPhone & iPad

Install through Apple TestFlight. This provides the version approved for external testers; newer source changes are not included automatically.

Open in TestFlight
In preparationMac

The native app with bundled runtime is in local testing. A publicly installable, signed and verified Mac release is not available yet.

No public download yet
In preparationWindows

The desktop app is in testing. A publicly verified installer is not available yet.

No public download yet

Now

ApprovalsWhat waits for your credential
ThreadsWhat runs alongside
IntentionWhat's up next

Understand

BrainMemory, identity, conversations
ScreenWhat could be read on the computer

Connect

ComputersConnect and control computers
ChannelsTelegram, WhatsApp — reach Hunch
NodesDevices in the Hunch network

Tools

Skills & MCPsWhat the runtime can and may do
AppsWhat Hunch knows about apps
GitHubFind and install agent repositories
Hunch ID card with HID
Silent screen recording of the Hunch app. It scrolls through the Hub to Nodes and Tools, then shows Brain, Profile and Settings.
Brain — personal memory
HubFour quiet groups, the same 11 routes. The image shows the opening view; the video continues to Nodes and Tools, while the register above lists every route.
Hunch hub
BrainMemories, tasks, goals and conversations — locally on the device.
Brain overview
Any modelObserver and Executive freely on any provider — cloud, local or on-device.
Model selection
ID cardSecure Enclave identity with HID — shareable via QR, no app needed on the other side.
ID card
VoiceTalk to Hunch in a visible session. Experimental, no always-on microphone.
Voice mode
ChannelsTelegram and WhatsApp as entry points — execution runs on the runtime.
Channels

Work Live Activity in the Dynamic Island

WorkingThread running · progress & runtime, deep link back to the task
Waiting for youApproval required — tool, target and arguments visible

States: thinking, working, waiting, done, failed.Today— Voice uses background audio only in an already active session; the app does not start spontaneously in the background.

Local-first & control

Not just “privacy first” as a slogan — this is how control works.

origin documentThe ID card is the genetic origin of the interface: security paper, engraving, guilloché, pearl, foil, physical edge, owner-bound seal.
format1.586 : 1 · max. 340 pt · radius 28 · tilt max. 7°, reduced motion fades instead of tilting.
honestLooks are no proof: guilloché signals identity — verification comes from signature, authority, provenance and audit.
specimenAnonymised data, no external identity. A real HID is derived from your device's Secure Enclave key.
iPhoneapp · brain local
Mac / PC / VPSruntime · memory · tools
no relay · no hosting · no service in the middle

Local & your hardware

Personal memory lives locally by default. The runtime runs on your own hardware.

Sensors opt-in

Sensors are activated explicitly, imports are opt-in. No “Hunch reads everything”.

Isolation & authority

Guests are isolated, critical tools need authority, external agents only get limited interfaces.

Foreign content = distrust

Foreign content is treated as potentially untrustworthy and handled as a quote, never as a command.

Visible & deletable

Memories, preferences and learned skills are visible, controllable and deletable.

Mood is a signal

Tone of voice is a heuristic signal, not real emotion recognition.

Honest status

What runs today, what is experimental, what is coming.

Nothing here is presented as shipped that isn't. “Today” means: present and tested in the code. Which version you can install is listed in the delivery status under Apps.

today

Present and tested in current code; runtime-dependent claims are additionally proven against the running runtime.

experimental

Prototype or on one surface only — marked visibly.

next

Planned, not yet built or not yet shipped.

vision

Towards 2.0 — explicitly not shipped.

Today

  • Persistent personal memory
  • Observer & Executive
  • Threads
  • Provider choice, local & cloud models
  • SKILL.md, Python skills, MCP, REST, harnesses
  • Approvals, risk classes, budgets, audit
  • Secure Enclave identity, guest isolation
  • Runtime, app, CLI, HTTP & WebSocket
  • Brain sync, computers, nodes, channels
  • Work Live Activity, Telegram/WhatsApp
  • Intentions with state & next step (runtime, app, agent tools, sync)
  • Automatic condensation (every 6 h, logged)
  • Skills learned from routines
  • Memory levels global / intention / thread — store, search, move, sync (runtime, app, MCP, CLI)
  • Broader providers: OpenRouter, Together, Fireworks, Cerebras, Perplexity, Amazon Bedrock
  • Logbook: every action with reasoning, outcome and way back (runtime & app)
  • Way back for everything recoverable — memories, tasks, intentions & their changes, levels, learned skills, files; shell/agent honestly marked as not reversible
  • Device & action boundaries: every device, every channel up to its risk class (read only · read and remember · everything)
  • Knowledge graph & Pattern of Life: topics, people, places, files, intentions, threads as a graph; hour profile, transitions, cadences (runtime, app → Brain → Graph)
  • App/window/browser/file context as sensors of the computer (switchable, quiet hours; screen OCR via macOS Vision, behind the screen-recording permission), rewind via MCP, transcript import (ChatGPT/Claude export, text)
  • Proactive predictions & nudges: prediction nudges in the heartbeat, delivered to the active surface
  • Presence routing: the computer knows which surface is active right now and delivers there
  • Intentions on desktop surfaces: hunch vorhaben, web UI, cockpit
  • Agent vault: logins, payment methods, passkeys from the wallet onto the computer — every use through the Authority Gate, payment only with purpose and amount, the secret never in the log
  • Intention runner: set a goal, plan of steps, steps as threads, across days and restarts; failure → replan once, then pause and ask; questions only where the gate requires them
  • Triggers: time, file change, failed build, channel message → job or thread; resume paused threads
  • Shared history + presence: app, runtime, CLI, web and Telegram mirror conversations; see external conversations in the Brain and continue with their last 30 lines; delivery goes to the active or reachable surface

Experimental

  • Hunch ID: in-app owner confirmation with Face ID or Touch ID; full physical-device acceptance testing remains open
  • Apple Watch: the signed iOS archive 200 includes the Watch app and complication. Installation and approvals on a physical Watch have not yet passed acceptance testing
  • Approve/decline from a push notification: code and APNs acceptance are proven; the corrected button on build ≥118 awaits end-to-end confirmation
  • Prediction push in daily use: generation and presence routing run, but actual push delivery has not yet been confirmed
  • Companion: main view in TestFlight 234, classic chat optional under Settings → View. The same conversation and draft; nine original animals in the catalogue, three experimental bird models available. ChatGPT model access and transmitted CLI context remain bounded without deleting history. Input and error messages with the keyboard open and large accessibility text have been improved and simulator-tested; the specific device error remains unverified. Experimental on iOS 26+: explicitly speak a Gemini command and choose “Done”; only Apple's locally finalized text goes to the connected agent. No automatic recording or speech-model downloads; actual Voice device tests remain open
  • Voice: a conversation you start visibly, with listening and answers. Background conversation and computer requests through to the actual result still require physical-device acceptance; no hidden always-on microphone
  • Approval levels: you decide how much Hunch may do without asking, from Sandbox up to a deliberately chosen Unsafe mode. Present in runtime and apps; the effect per surface has not been fully accepted yet
  • Daily report and health: selectable sources and purposes, evening report only after activation. The native health view shows imported measurements, provenance, a personal baseline and data gaps. Source/device acceptance remains open; no continuous recording or diagnosis
  • Hunch keyboard (iPhone): its own keyboard with local word suggestions. Fast typing has not yet been accepted as error-free

Next

  • Planned starting model: no Hunch usage fee. Your own provider costs, hardware and energy remain separate. Paid extensions have not been decided; this is not a price or purchase offer
  • Mac and Windows app as a public download once signing and device testing have passed
  • Daily companion that picks the time of the evening report from your day instead of a fixed hour

Vision 2.0

  • A lasting counterpart across models, devices and years

Deliberately not planned: a hosted relay service. Reachability away from home runs over a private network of your choice (VPN/mesh) plus TLS — your devices, no third party in the middle.

knows you.understands the moment.acts on your behalf.official line · brand system 2026.03

Hunch is not the model you talk to today.

It is the agent that gets to know you over years, understands where you are heading, knows how you want to work with it — and acts at the right moment within your rules.

Try the beta in TestFlight
credentialowner bound · secure enclavescopeglobal · project · threadauthorityapproval before effect · undo · auditstatustoday / experimental / vision — per product truth

Questions: schenkei@hunchagent.io · no newsletter, no trackers on this page.