Market research · Agent ↔ human communication

The market hasthe wires.Nobody owns the handoff.

AI agents can already send a push, email, Slack message, SMS, or phone call. What remains fragmented is the trustworthy loop: reach the right person, collect a typed decision, resume safely, and prove what happened.

20+ products mapped45+ public sourcesReviews + adoption proxiesPrimary research cutoff: 17 Jul 2026

Live relay tracehandoff incomplete

Agent requests a decision

“Approve refund of €840?”

running

identity · policy · durable state

Person responds anywhere

Approve · Reject · Edit · Explain

waiting

WATCH
PUSH
EMAIL
SMS
VOICE

The missing product is not a notification API. It is a durable, secure decision exchange that happens to use notifications.

01 · Executive verdict

Yes, products exist. No, the category is not settled.

There are strong point solutions and fast-moving platform features, but no clearly dominant horizontal product combines arbitrary-agent support, native phone/watch UX, rich input, multichannel escalation, durable resume semantics, and an audit-grade decision record.

8/10

Need intensity

Agents increasingly run asynchronously, while governance and trust lag adoption.

6/10

Whitespace

The complete horizontal loop is open, but several young products are chasing it.

8/10

Timing

MCP, mobile remote control, and first-class HITL patterns are converging now.

5/10

Defensibility

Delivery is commoditized; defensibility must come from workflow state and trust.

Demand is real

LangChain’s 2026 survey says production agent use rose from 51% the prior year, while Deloitte found only 21% of organizations had mature agent governance.41 42

The obvious MVP exists

GetApproved, PausePoint, Impri, HITL.sh, and legacy HumanLayer already sell variants of “pause, ping a human, resume.” Most have little independent adoption evidence.

Distribution is the fight

Claude, OpenAI, Happy, and Omnara own the best UX where they own the agent session. A horizontal entrant must win where agents and channels are heterogeneous.

How to read the scores. They are an informed product judgment, not a statistical market model. Popularity is based on observable proxies—store ratings, review volume, GitHub activity, public customer claims, and community evidence. Vendor claims are labeled; absence of evidence is not treated as evidence of failure.

02 · Market map

Five markets are colliding around one job.

Competitors approach the problem from different directions. That matters: the same “send an approval” demo can hide very different strengths in session control, notification delivery, workflow design, or communication identity.

Direct HITL

Closest to the idea. A request record, human response surface, and callback to the agent.

GetApproved · PausePoint · Impri · HITL.sh · HumanLayer legacy

Agent control

Best end-user experience. Mobile access to a running agent, with push, approvals, diffs, and steering.

Happy · Omnara · Claude Remote · OpenAI Remote

Workflow suites

Broadest business reach. Add approval nodes and route them through existing workplace channels.

n8n · Zapier · Power Automate

Delivery infra

Channels are solved. Routing, preferences, templates, failover, analytics, push, email, SMS, and chat.

Courier · Novu · Knock · SuprSend · OneSignal

Primitives

Cheap building blocks. Simple push, actionable notifications, email inboxes, phone numbers, and voice.

ntfy · Pushover · Pushcut · AgentMail · AgentPhone · Twilio

Category boundary: “PagerDuty for agent decisions.”

PagerDuty does not merely send a push; it owns incident state, acknowledgement, escalation, on-call identity, and audit. The analogous opportunity is to own the lifecycle of an agent’s human dependency—not merely the transport that announces it.

03 · Competitors

Who is already doing it—and how well?

The cards separate capability from traction. Use the filters to compare direct entrants with platform, workflow, delivery, and channel-native alternatives.

Happy

Agent control

Open-source mobile remote control for Claude Code, Codex, and related coding agents: push, permission handling, task completion, session steering, diffs, and end-to-end encryption.

22.7kGitHub stars

4.9 / 962iOS rating

10k+Android installs

Popularity: The strongest independent adoption signal among dedicated mobile agent controllers; 2.7k+ Google Play reviews supplement the store and GitHub evidence.1 2 3
Happiness: Very high ratings. Reviewers praise the intuitive mobile context and response suggestions; public issue volume also shows the cost of wrapping fast-changing agent CLIs.
UX: Excellent for supported coding agents: install CLI, scan QR, use the phone. Not a neutral human endpoint for arbitrary business agents.

Implication: validates the mobile “agent needs me” experience, but leaves the horizontal market open.

Omnara

Agent control

Commercial mobile/web control plane for coding agents, with push, approvals, diffs, voice interaction, local/cloud handoff, and multi-agent orchestration.

10k+Android installs

4.7 / 42Android rating

2.7kLegacy repo stars

Popularity: YC S25, a #1 Product Hunt launch with 445 points, and a claimed 250k agent interactions in its first launch week; present public review volume is still modest.4 5
Happiness: Generally positive and praised for QR onboarding and mobile UI. Reviews also report missing Android alerts, stale/waiting state, and occasional endless sends.4 6
UX: Polished and focused, but the original wrapper was archived after CLI changes became hard to maintain; Omnara rebuilt on an integrated agent SDK.5

Implication: strong product sense; also a warning against coupling the business to one agent’s terminal protocol.

Claude Remote Control

Platform native

Anthropic’s official phone, tablet, and browser control for a local Claude Code session, including synced conversation, decisions, and mobile push.

OfficialBundled channel

iOS + AndroidMobile surfaces

PreviewFeature maturity

Popularity: Feature-specific usage is not public, but distribution through the main Claude app is a formidable advantage.
Happiness: Community reaction is positive around replacing tmux/SSH workarounds, with reports of rollout, session, and memory limitations typical of a preview.7 8
UX: Native and low-friction. Push is intentionally simple: Claude decides when to notify; beyond an on/off toggle, there is no per-event configuration.7

Implication: impossible to beat inside Claude Code on convenience; compete across vendors and business workflows.

OpenAI Remote

Platform native

Official remote access from the ChatGPT mobile app to desktop-hosted work: start or continue tasks, approve actions, inspect output, and receive attention/completion notifications.

OfficialBundled channel

MobileApproval surface

ExperimentalCLI remote control

Popularity: Specific Remote usage is not disclosed. Distribution through ChatGPT and Codex makes it a structural threat to generic coding-agent companions.
Happiness: Too new for reliable feature-specific sentiment. The product covers the desired job when both the agent and person live inside OpenAI’s environment.
UX: QR pairing, mobile approvals and follow-ups, output/diff review, and notifications. It remains a first-party surface, not infrastructure another agent can call.9 10

Implication: platform bundling raises the urgency of owning a vendor-neutral identity and decision record.

GetApproved

Direct HITL

A close conceptual match: push to a mobile interface, with confirm, multiple choice, text, numbers, forms, webhooks, team members, audit logs, Slack, and email.

$29/moPro list price

50Free requests

None foundIndependent reviews

Popularity: No verifiable store listing, GitHub community, review corpus, or customer-scale disclosure was found. Website testimonials are not independent evidence.
Happiness: Unknown. The feature set is persuasive, but public user evidence is insufficient.
UX: Claims two-minute setup and sub-ten-second responses. Strong input breadth; integration appears webhook/REST-led rather than a deeply specified agent lifecycle.11

Implication: proves the exact pitch is in market; differentiation cannot be “push + forms” alone.

PausePoint

Direct HITL

One API/MCP call creates an SMS approval, hosted response page, timeout, webhook/polling result, and append-only audit log; email and Slack arrive on paid tiers.

$29/mo500 pauses

50Free pauses

None foundIndependent reviews

Popularity: Very early. No independent adoption or satisfaction corpus was found.
Happiness: Unknown. The no-app SMS link reduces onboarding friction, but rich/freeform input is less visible than GetApproved.
UX: Good developer story: no Twilio account, explicit safe timeout defaults, signed webhooks, and one-time links. Phone-number-first UX trades speed for messaging compliance burden.12

Implication: “no app required” is a strong alternative onboarding path worth testing.

Impri

Direct HITL · OSS

Open-source approval inbox with REST, MCP, Python/TypeScript SDKs, web push, Slack, Telegram, Discord, rules, bulk decisions, edit-before-approve, and audit logs.

1GitHub star

Jul 2026Public launch

Self-hostDeployment option

Popularity: Newly launched with essentially no traction signal yet; public Reddit threads have only a handful of votes/comments.13 14
Happiness: Too early to measure. Early feedback correctly pushes beyond approve/reject toward risk, expected impact, rollback, and execution receipts.
UX: Broad builder surface and good structural gating. Mobile is a PWA/web-push experience rather than a proven native phone/watch product.

Implication: the open-source baseline is already rising; hosted trust and mobile UX must be materially better.

HumanLayer

Direct HITL · legacy

The original well-known horizontal pitch: SDK decorators for approvals and human-as-tool conversations across Slack, email, SMS, Discord, escalations, and timeouts.

11.1kCurrent repo stars

YC F24Launch signal

LegacySDK status

Popularity: The GitHub repo is popular, but its current identity is coding-agent tooling and consulting; the original HITL SDK is explicitly labeled legacy, so stars cannot be read as approval-product usage.15
Happiness: Early launch testimonials praised Slack-based feedback and framework independence, but current satisfaction for the old service is not observable.16
UX: Strong API concept and channel breadth; weak continuity as a standalone commercial category.

Implication: validates the problem and warns that HITL alone may not sustain a company without a larger control-plane wedge.

n8n

Workflow suite

First-class gating of AI tool calls with approve/deny through Slack, Discord, Telegram, Teams, Gmail, WhatsApp, Google Chat, Outlook, or n8n Chat.

196.7kGitHub stars

4.7 / 272G2 rating

9Approval surfaces

Popularity: By far the largest open builder ecosystem in this set; GitHub and G2 both show meaningful adoption.17 18
Happiness: Users praise flexibility, self-hosting, integrations, and the visual builder; recurring drawbacks are learning curve, missing features, and interface complexity.
UX: Very capable for a builder already in n8n. End-user review is limited to approve/deny and channel-specific rendering; users call for richer, multimodal review context.19

Implication: n8n is both a competitor and the best early distribution partner.

Zapier HITL

Workflow suite

A built-in step pauses a Zap for approval or data collection, allows submitted content to be edited, and notifies through email, Slack, or a secondary Zap.

8,000+App ecosystem claim

Email + SlackNative review

MatureAutomation brand

Popularity: Zapier’s installed base and integration catalog make it a serious substitute for nontechnical teams, though HITL-specific usage is not disclosed.
Happiness: The approval setup is approachable. Community questions show previous confusion about waiting for Slack responses and resuming the same workflow.
UX: Strong no-code builder UX, but mobile intervention routes to a web run and Zapier lacks a strong mobile command-center experience.20 21

Implication: do not compete on workflow authoring; make the human decision endpoint Zapier can call.

Novu

Delivery infra

Open-source notification workflows across in-app, email, SMS, push, and chat, with subscriber preferences, inbox components, an MCP server, and agent toolkit.

39.3kGitHub stars

4.6 / 22G2 rating

55+Provider integrations

Popularity: Strong developer awareness and active open-source development; review volume is much smaller than OneSignal, Knock, or SuprSend.22 23
Happiness: Reviews praise setup, support, consolidated channels, and developer experience; complaints include beta features, unclear docs, subscriber flakiness, short retention, and missing template history.
UX: Good for building notification infrastructure. MCP mainly configures and triggers workflows; it does not provide a complete human decision/resume contract.

Implication: credible open-source delivery layer to integrate or learn from—not the product to duplicate.

Courier

Delivery infra

Unified API and MCP tools for email, SMS, push, inbox, Slack, Teams, WhatsApp, Discord, webhooks, orchestration, preferences, and delivery observability.

4.5 / 59G2 rating

9+Channel families

35+MCP tools claimed

Popularity: G2 lists thousands of companies and recognizable customers; review count is credible but not category-leading.24
Happiness: Users praise ease, flexibility, APIs, and support. Criticism includes occasional slowness and missing social integrations.
UX: Broadest B2B channel mix and increasingly AI-friendly tooling. Like Novu, it sends and observes; durable structured human response remains application logic.

Implication: leading candidate to rent delivery breadth while owning the decision state above it.

OneSignal

Delivery infra

Large push-first engagement platform spanning mobile/web push, email, SMS/RCS, in-app messages, Live Activities, journeys, and a beta MCP server.

1,211G2 reviews

1m+Live apps · vendor claim

12bn+Daily sends · vendor claim

Popularity: Clear category-scale incumbent. OneSignal claims 2m+ developers/marketers and one in five new mobile apps; these are vendor figures, but the independent review volume is also strong.25 26
Happiness: Reviews consistently praise ease, dashboard, setup, and free tier; advanced features, cost, missing features, and notification issues recur as drawbacks.
UX: Excellent at reaching an app’s users and managing campaigns. MCP manages messaging resources; it is not a human approval state machine.27

Implication: building generic push infrastructure would be strategically indefensible.

Knock & SuprSend

Delivery infra

Developer-focused notification platforms with workflow engines, templates, preferences, in-app components, observability, and broad channel/provider support.

4.6 / 144Knock · G2

4.8 / 127SuprSend · G2

MCPSuprSend native

Popularity: Both have credible product-team adoption and more review evidence than the direct HITL startups.28
Happiness: Knock reviews praise powerful workflows, usability, developer/product collaboration, and support; complex documentation and edge limits surface. SuprSend scores especially well on usability and requirements.
UX: Strong builder/admin UX. SuprSend’s MCP can build and trigger notifications; human approval still needs a response mechanism and durable state outside the delivery step.29

Implication: the delivery market is healthy, competitive, and partner-rich.

ntfy

Channel primitive

Open-source HTTP pub/sub notifications to mobile and desktop, plus action buttons, email forwarding, phone calls through Twilio, self-hosting, and community MCP servers.

31.9kGitHub stars

100k+Android installs

4.9 / 1.6kAndroid rating

Popularity: Large, active developer/self-host community with substantial store adoption.30 31
Happiness: Very positive on simplicity, reliability, self-hosting, and immediate delivery. iOS reviews ask for two-way publishing, filtering, image fixes, and better notification control.32
UX: “Pick a topic, POST a message” is hard to beat for builders. HTTP action buttons can approximate approvals, but identity, forms, audit, and agent resume are DIY.33

Implication: the best zero-friction technical substitute—and an excellent prototype transport.

Pushover / Pushcut

Channel primitive

Mature personal push tools. Pushover offers reliable phone/watch delivery and API/email gateways; Pushcut offers webhook-triggered, actionable iOS/watch notifications.

4.8 / 3.4kPushover iOS

$4.99Pushover one-time

WatchNative actions

Popularity: Pushover has years of credible operational use and one of the strongest satisfaction signals in the channel-primitives set.34
Happiness: Reviews emphasize reliability and setup ease. One long-standing request is revealing: approve/deny responses with backend postback, beyond simple acknowledgement.
UX: Very good notification UX, including Apple Watch. Pushcut proves web requests can run from watch actions, but both require the builder to own the approval lifecycle.35

Implication: wearable approval is technically routine; trusted decision semantics are the value.

AgentMail / AgentPhone

Agent identity

Agent-native communication identities: AgentMail gives agents API-first inboxes; AgentPhone provisions numbers for SMS, iMessage, and voice calls. Both expose agent-friendly APIs/MCP.

YC S25AgentMail

$6mFunding · founder claim

5.0 / 4AgentMail · PH

Popularity: AgentMail has credible funding and launch interest but a small independent review corpus. AgentPhone is YC-backed and claims use by teams around major agent frameworks.36 37
Happiness: Too early for reliable satisfaction conclusions. The category is clearly active: agent-native email and phone identity are becoming products of their own.
UX: Strong for agents communicating with external people, not for securely obtaining an owner’s approval and resuming a suspended run.

Implication: partner/channel providers; do not become an email or telephony carrier.

04 · Popularity & sentiment

The happiest users are in mature push tools and focused mobile controls.

The closest horizontal HITL products are too early to evaluate. The strongest evidence comes from adjacent tools that solve one part extremely well.

#	Product	Best public popularity signal	User sentiment	Usability read	Evidence caveat
01	n8n	196.7k GitHub stars; 272 G2 reviews	4.7 / 5	Flexible and visual; steep for nontechnical users	Suite-level, not HITL-only data
02	OneSignal	1,211 G2 reviews; vendor claims 1m+ live apps	4.7 / 5	Easy core setup; advanced features can be opaque	Engagement platform, not approval product
03	ntfy	31.9k GitHub stars; 100k+ Android installs	4.9 / 5	Minute-level builder setup; incomplete two-way UX	Technical/self-host audience
04	Happy	22.7k GitHub stars; 962 iOS + 2.7k Android ratings	4.9 / 5	Very strong phone UX for supported coding agents	Coding vertical only
05	Pushover	3.4k iOS ratings; long operating history	4.8 / 5	Reliable, fast, watch-friendly	Notification/acknowledgement, not decisions
06	Knock	144 G2 reviews	4.6 / 5	Excellent builder/product collaboration	Some reviews seller-incentivized
07	SuprSend	127 G2 reviews	4.8 / 5	High reported usability and requirement fit	Delivery platform
08	Courier	59 G2 reviews; “thousands” of customers claimed	4.5 / 5	Easy and flexible; occasional slowness	Vendor customer count
09	Novu	39.3k GitHub stars; 22 G2 reviews	4.6 / 5	Good DX; beta/retention/admin rough edges	Stars ≠ managed-cloud usage
10	Omnara	10k+ Android installs; 42 reviews	4.7 / 5	Polished; bugs and narrow agent coverage reported	Legacy repository is archived
11	GetApproved	No independent adoption signal found	Unknown	Compelling claimed UX	Website claims only
12	PausePoint	No independent adoption signal found	Unknown	Very simple no-app SMS path	Very early
13	Impri	1 GitHub star at research cutoff	Too early	Broad open-source feature set	Launched days before cutoff

Interpretation: rankings reflect strength of public evidence, not product quality or revenue.GitHub counts captured 17 Jul 2026.

Users love “it just pings me.”ntfy and Pushover reviews consistently reward immediate delivery, low setup, and reliability.

Users dislike broken state.Stale diffs, contradictory waiting status, missing push, and lost context are more damaging than missing features.

Rich context matters.n8n community feedback asks for media, warnings, and better review context—not merely another approve button.

05 · Capability matrix

No product owns the full row.

Native means the capability is part of the core product. Partial means it is limited, community-built, workflow-configurable, or possible through another provider.

Capability	Happy	GetApproved	PausePoint	Impri	n8n	Courier / Novu	ntfy
Arbitrary agent via MCP/API	◐	◐	●	●	◐	●	◐
Native phone push	●	●	○	◐	◐	◐	●
Watch actions	◐	○	○	○	◐	◐	◐
Email / workplace chat	○	●	●	●	●	●	◐
SMS / phone call escalation	○	○	◐	○	◐	◐	◐
Approve / reject	●	●	●	●	●	○	◐
Choice / text / typed forms	◐	●	◐	◐	◐	○	○
Durable wait + agent resume	●	◐	●	●	●	○	○
Policy, routing, escalation	○	◐	◐	◐	◐	●	○
Decision + execution audit	◐	◐	◐	◐	◐	◐	○
Meaningful public adoption	●	⚗	⚗	⚗	●	●	●

● native◐ partial / buildable○ not found⚗ too early to judge

06 · The whitespace

The gap is a protocol plus an opinionated human surface.

Six missing pieces recur across the market. Together they form a product boundary that neither notification vendors nor agent runtimes naturally own.

GAP 01

A universal human endpoint

One person identity reachable by policy across app push, watch, email, Slack/Teams, SMS, or voice—independent of which agent vendor initiated the request.

GAP 02

Typed, portable response semantics

Approve/reject is insufficient. Agents need JSON-schema input: choice, text, number, date, file, form, edit, delegation, and “ask someone else.”

GAP 03

A durable resume contract

Long MCP calls time out. A request must survive process restarts and return through polling, webhook, event stream, or a framework adapter that checkpoints the agent.

GAP 04

Approval integrity

The person must approve an exact operation—not a summary the agent can later reinterpret. Bind the decision to an immutable payload hash, scope, expiry, and actor.

GAP 05

Outcome receipts

An approval record is incomplete without what executed afterward. Store proposal → decision → tool result → affected resource IDs → rollback/undo reference.

GAP 06

Attention policy

Quiet hours, urgency, batching, deduplication, first-response-wins, escalation, on-call routing, and channel preferences prevent “human in the loop” from becoming spam.

The moat is the ledger, not the ping.

Every channel vendor can copy an MCP send tool. Fewer can become the system of record for who authorized which irreversible agent action, under which policy, and what actually happened next.

Accumulated decision and policy history
Framework adapters that resume agents reliably
Recipient identity, teams, roles, and escalation graphs
Trusted native mobile/watch review UX
Compliance export and execution receipts

07 · What to build

Start with agent builders who already feel the pain.

The first buyer should not be every ChatGPT user. Target technical teams and automation agencies running long-lived, heterogeneous workflows whose customers or colleagues must approve real-world actions.

Positioning

The decision inbox for every agent.

“Give any agent a safe way to reach the right person, collect a structured decision, and resume—with a record you can trust.”

Do not position as

another notification API, another mobile coding client, or a generic workflow builder.

Best initial wedge

n8n / Make / OpenClaw automation buildersThey assemble HITL glue repeatedly, serve many end-users, and already use webhooks/MCP.
Internal operations agentsOutbound email, CRM changes, refunds, procurement, publishing, and production actions create obvious approval moments.
Agent platform teamsThey need a white-label human endpoint across customers without building apps, preferences, and escalations themselves.

Phase 1 · Prove response

Decision core

Remote MCP + REST API
Notify, ask, approve tools
Mobile web/PWA + email fallback
Choice, text, number, form
Polling, webhook, expiry, cancel
Immutable decision record

Phase 2 · Own attention

Native human UX

iOS + Android push
Watch approve/reject actions
Slack, Teams, Telegram, WhatsApp
Quiet hours and preferences
Teams, roles, delegation
Escalation and reminders

Phase 3 · Enterprise trust

Decision control plane

Execution receipts and rollback
Policies and conditional approval
Multi-party/quorum decisions
SSO, SCIM, audit export
SMS and voice escalation
Self-host / regional residency

Keep the MCP vocabulary tiny.

Agent-facing tools should describe intent, not channels. Routing belongs to the person’s policy; the agent should not decide to wake someone by phone unless explicitly authorized.

notify: Inform; no response required.
ask: Request structured input described by JSON Schema.
request_approval: Bind a human decision to an exact proposed action.
get_request: Read status/result without keeping an MCP call open.
cancel_request: Close stale work and retract outstanding prompts.

request_approval({
  recipient: "role:finance-on-call",
  operation_id: "refund:ord_1842",
  summary: "Refund €840 to Ada",
  payload_hash: "sha256:9b1…",
  risk: "irreversible",
  expires_in: "30m",
  on_timeout: "reject"
})

→ { request_id: "req_7f2",
    status: "pending" }

Recommended system boundary

Agents & workflowsMCP, REST, OpenAI Agents, Claude SDK, LangGraph, n8n, Zapier.

→

Your decision control planeRequest state, schema, identity, policies, escalation, audit, payload binding, resume adapters.

→

Rented deliveryAPNs/FCM plus Courier/Novu/OneSignal; Twilio or AgentPhone later for SMS/voice.

Pricing must reflect decisions, not messages.

Direct entrants anchor the market around free trials and roughly $29/month individual/pro tiers. A credible entry model is free for experimentation, a low-friction individual plan, and team pricing based on active approvers plus decision volume; pass through SMS/voice costs.

Free · 50/moOne approver, push/web/email, community support.

Builder · $19/mo1,000 decisions, all input types, webhooks, basic audit.

Team · $99/moFive approvers, routing, escalation, integrations, exports.

08 · Risks

The biggest risk is being useful but unnecessary as a company.

A thin HITL API can be reproduced with a webhook, database row, and message. The product must own enough workflow state and trust to earn recurring value.

High

Platform bundling

Anthropic and OpenAI now provide remote mobile approval in their own products. Mitigation: serve cross-vendor agents, business workflows, teams, and external approvers.

High

Commodity delivery

Notification MCP servers are already native in OneSignal, Novu, Courier, and SuprSend. Mitigation: never make “send” the core value or core architecture.

High

Approval theater

A person may approve a friendly summary while the agent executes a different payload. Mitigation: immutable operation IDs, hashes, scope, expiry, and execution receipts.

Medium

Human bottlenecks

Over-gating produces fatigue and slow workflows. Mitigation: policy thresholds, digests, confidence/risk routing, delegation, and measured auto-approval.

Medium

Channel compliance

US app-to-person SMS requires registration and verifiable consent; AI-generated voice calls generally require prior express consent under the TCPA.38 39 Mitigation: ship push/email first and outsource carrier compliance.

Medium

Resume fragmentation

Every agent framework checkpoints differently, and long MCP calls are brittle. Mitigation: asynchronous request IDs plus framework-specific resume adapters.

The minimum trustworthy approval record

An audit trail should record the whole chain, not just “Vegard clicked approve.”

1 · ProposalAgent, run, tool, exact payload hash, risk, intended target.

2 · DecisionHuman identity, factors, scope, channel, timestamp, comment.

3 · ExecutionWhat tool actually ran, result, resource IDs, errors, retries.

4 · RecoveryRollback reference, undo window, compensation status.

Validate before building the channel stack

Run a concierge pilot on top of PWA/web push + email and one delivery provider. Interview 12–15 automation builders, then integrate three real workflows where the human is not the developer who configured the agent.

Activation: first real decision completed in under 10 minutes from signup.
Response: median human response under 5 minutes for urgent requests.
Breadth: at least 25% of requests need more than yes/no.
Channel value: at least 30% of responses occur outside the builder’s default chat tool.
Retention: five teams still send weekly decisions after four weeks.
Willingness to pay: three teams pay before native phone apps or voice exist.

09 · Decision

Proceed—if the company is about trusted decisions, not multichannel notifications.

The market validates every component: people want mobile agent control; builders pay for notification infrastructure; businesses need HITL; and MCP makes distribution portable. The opening is the connective tissue those products leave behind.

Build first: universal decision inbox + MCP/API + push/email + typed input + durable state + audit. · Add later: Slack/Teams, watch, SMS, and phone escalation.

10 · Sources

Evidence library

Primary product documentation is used for capabilities. GitHub, app stores, G2, Product Hunt, and community threads provide adoption and sentiment proxies. All links were checked during research on 17 July 2026.

Mobile agent control

01Happy GitHub repository — stars, activity, open-source implementation.
02Happy on the App Store — rating, review volume, product UX.
03Happy on Google Play — installs and review volume.
04Omnara on Google Play — installs, rating, positive and negative reviews.
05Omnara legacy repository — archive/pivot notice and prior traction.
06Omnara on the App Store — rating and reported state/diff issues.
07Claude Code Remote Control docs — mobile control, notification behavior, limitations.
08Claude Code community thread — early user reaction and rollout friction.
09OpenAI Remote connections — mobile tasks, approvals, review, notifications.
10OpenAI Notifications — desktop, web, push, email, SMS, and CLI surfaces.

Direct HITL products

11GetApproved — input types, push workflow, integrations, pricing.
12PausePoint — SMS approvals, timeouts, audit, pricing, MCP claim.
13Impri documentation — MCP, SDKs, channels, inbox, audit.
14Impri GitHub repository — current public adoption/activity signal.
15HumanLayer repository — current positioning and legacy SDK status.
16HumanLayer YC launch — original channels, pitch, early testimonials.
—HITL.sh — mobile approval product and public pricing.
—HITL.sh roadmap — MCP listed as in development.

Workflow & delivery platforms

17n8n GitHub repository — popularity and project activity.
18n8n G2 reviews — rating, review count, usability themes.
19n8n HITL docs — approve/deny behavior and nine review channels.
20Zapier HITL guide — approval, data collection, edit, notification flow.
21TechRadar automation comparison — Zapier mobile and complexity assessment.
22Novu GitHub repository — stars and project activity.
23Novu G2 reviews — rating and satisfaction themes.
24Courier G2 reviews — rating, review count, satisfaction themes.
25OneSignal G2 reviews — rating, volume, ease/cost themes.
26OneSignal scale claims — businesses, daily sends, apps, and monthly active users.
27OneSignal MCP docs — beta scope and supported actions.
28Knock G2 reviews — rating, volume, usability, limitations.
29SuprSend MCP — build, trigger, query, and debug notifications.

Channel primitives

30ntfy GitHub repository — stars, releases, activity, feature overview.
31ntfy on Google Play — installs, reviews, reliability feedback.
32ntfy on the App Store — rating and two-way/filter/image feedback.
33ntfy action-button docs — HTTP callback actions from notifications.
34Pushover on the App Store — rating, watch support, approval demand.
35Pushcut notifications docs — webhook actions and Apple Watch behavior.
36AgentMail on Product Hunt — launch, reviews, positioning.
37AgentPhone YC profile — product and early ecosystem claims.
—Twilio MCP docs — agent access to messaging and voice APIs.
—AgentPhone docs — SMS, calls, voice, webhooks, MCP.

Compliance & device behavior

38Twilio A2P 10DLC docs — registration, consent, sender identity.
39FCC declaratory ruling — AI voice calls and TCPA consent.
40Apple watchOS notification actions — watch buttons and suggested text responses.
—Pushover pricing — one-time individual price and team tier.
—ntfy call/email configuration — Twilio phone calls and SMTP forwarding.

Market demand & governance

41LangChain State of Agent Engineering 2026 — 1,300-person survey and production-use trend.
42Deloitte: agents scaling faster than guardrails — 3,235 leaders; governance gap.
—Grant Thornton AI Impact Survey 2026 — high-stakes human review and autonomy limits.
—KPMG Q2 2026 AI Pulse — approval processes, agent deployment, oversight skills.
—How AI agents are used: 177k MCP tools — agent-tool ecosystem scale and concentration.
—Human Tool paper — MCP-style abstraction that exposes people as callable tools.

Limitations. Private revenue, active-user, retention, and request-volume data were generally unavailable. App-store ratings can vary by country and over time. GitHub stars measure awareness, not production use. G2 includes vendor-solicited and sometimes incentivized reviews. Product claims and testimonials were not treated as independent user satisfaction evidence.

The market hasthe wires.Nobody owns the handoff.

Yes, products exist. No, the category is not settled.

Build the control plane, not another sender.

Need intensity

Whitespace

Timing

Defensibility

Demand is real

The obvious MVP exists

Distribution is the fight

Five markets are colliding around one job.

Category boundary: “PagerDuty for agent decisions.”

Who is already doing it—and how well?

The happiest users are in mature push tools and focused mobile controls.

No product owns the full row.

The gap is a protocol plus an opinionated human surface.

A universal human endpoint

Typed, portable response semantics

A durable resume contract

Approval integrity

Outcome receipts

Attention policy

The moat is the ledger, not the ping.

Start with agent builders who already feel the pain.

The decision inbox for every agent.

Best initial wedge

Decision core

Native human UX

Decision control plane

Keep the MCP vocabulary tiny.

Recommended system boundary

Pricing must reflect decisions, not messages.

The biggest risk is being useful but unnecessary as a company.

Platform bundling

Commodity delivery

Approval theater

Human bottlenecks

Channel compliance

Resume fragmentation

The minimum trustworthy approval record

Validate before building the channel stack

Proceed—if the company is about trusted decisions, not multichannel notifications.

Evidence library

Mobile agent control

Direct HITL products

Workflow & delivery platforms

Channel primitives

Compliance & device behavior

Market demand & governance