Evaluation record · poke

Poke

v1.x

The Interaction Company of California

Agentconsumermessaging-nativepersonal-assistantproprietary

Concerning

About This Agent

Consumer AI agent living entirely in messaging — iMessage/SMS, Telegram, WhatsApp — with no app to install. Routes each task to the best-fit model across providers; handles calendar, email, smart home, health, and purchases via recipes spanning 40+ integrations. Publicly launched March 2026; became the first third-party AI agent approved on Apple's Messages for Business (2026-06-04). Standing email/calendar access through a channel with no OS permission model is a novel risk surface.

Last Evaluated: July 9, 2026

Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability

task completion accuracy

Assessment of task completion across the recipe catalog from launch coverage and user reports; no public benchmarks exist

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Handles daily planning, calendar management, health tracking, smart home control, and photo edits over text; conversational disambiguation compensates for the low-bandwidth channel

mediumVerified: 2026-07-09

tool use reliability

Review of integration breadth and reliability of server-side tool execution

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Recipes integrate Gmail, Google Calendar, Outlook, Notion, Linear, Strava, Oura, Fitbit, Philips Hue, Sonos, GitHub, and 40+ other services

mediumVerified: 2026-07-09

multi step planning

Evaluation of trigger-based automations and proactive multi-step task chains

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Standing automations run on triggers such as every incoming email or real-time flight status, chaining multi-step actions without user prompting

mediumVerified: 2026-07-09

memory persistence

Review of long-lived conversational memory, the product's core design premise

Evidence

Poke — Poke is a persistent, always-on agent: one continuous conversation thread retains preferences, context, and connected-account state indefinitely

mediumVerified: 2026-07-09

error recovery

Assessment of failure visibility and recovery paths for background automations

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Failures surface as chat messages and the agent asks for clarification, but users cannot inspect or resume failed server-side automations

lowVerified: 2026-07-09

agent collaboration

Review of internal model/agent routing versus user-controllable collaboration

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Internally dispatches subtasks to the best-fit model per task across major providers and open-source options; no user-facing multi-agent composition

lowVerified: 2026-07-09

🛡️Security

tool sandboxing

Architecture review; cloud-side execution keeps the device safe but internal isolation is undisclosed

Evidence

Poke — All execution happens in Interaction's cloud rather than on the user's device; sandbox architecture and isolation between users' automations are not publicly documented

lowVerified: 2026-07-09

access control

Review of channel identity guarantees and OAuth grant model; the messaging channel bypasses OS permission prompts entirely

Evidence

TechCrunch - Apple approves Poke as the first AI agent on its Messages for Business platform — Apple's Messages for Business approval (2026-06-04) adds verified-sender identity and platform review on iMessage; per-service OAuth scopes exist, but SMS/Telegram channels carry no OS-level permission model and grants are standing rather than per-task

mediumVerified: 2026-07-09

prompt injection defense

Threat-surface analysis: email-triggered automations are a textbook indirect prompt-injection vector with no disclosed mitigations

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Automations that run on every incoming email make untrusted third-party content a standing trigger for an agent holding email, calendar, and purchase authority; no injection-hardening documentation is published

lowVerified: 2026-07-09

data isolation

Review of vendor security claims; positive design statements without independent verification

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Company states a multi-layered security model with regular penetration testing, that it cannot access integration token contents by default, and that log/analytics sharing is opt-in; TechCrunch notes no independent audit has verified these claims

lowVerified: 2026-07-09

open source transparency

Source availability assessment

Evidence

Composio - OpenPoke: an open-source alternative to poke.com — Poke's agent stack is fully closed; the community reverse-engineered an open-source clone (OpenPoke) precisely because internals are unpublished

highVerified: 2026-07-09

🔒Privacy & Compliance

data retention

Review of retention implications of a persistent standing agent against published policy

Evidence

Poke — The always-on conversation model retains message history, email/calendar context, and automation state server-side indefinitely by design; consumer privacy policy offers deletion on request but no enterprise-grade retention controls

lowVerified: 2026-07-09

gdpr compliance

Compliance posture assessment for a young consumer startup; no formal attestations found

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — 10-person Palo Alto startup with no published certifications (SOC 2, ISO 27001) or DPA; WhatsApp availability in the EU was still pending regulatory/platform constraints at launch

lowVerified: 2026-07-09

third party data sharing

Data flow analysis across model providers, messaging infrastructure, and carriers

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Best-fit model routing sends user content to multiple third-party model providers per task; messaging delivery relies on Linq's platform, and carrier SMS adds another processing party

mediumVerified: 2026-07-09

local deployment option

Deployment options assessment

Evidence

Poke — Cloud-only service reachable exclusively through third-party messaging channels; no self-hosted, on-premises, or offline option

highVerified: 2026-07-09

👁️Trust & Transparency

documentation quality

Documentation completeness review

Evidence

Poke — Consumer-grade site with recipe gallery and FAQs; no technical documentation of architecture, security model, or automation limits

mediumVerified: 2026-07-09

execution traceability

Review of action visibility; conversational reporting is not an audit trail

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Actions execute server-side and are reported back as chat messages; there is no action log, diff, or review step before the agent sends emails or makes purchases

mediumVerified: 2026-07-09

decision explainability

Assessment of conversational narration versus underlying decision opacity

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Agent narrates what it is doing and asks before consequential steps in normal use, but model-routing choices and automation internals are opaque

lowVerified: 2026-07-09

open source code

Open source assessment

Evidence

Composio - OpenPoke: an open-source alternative to poke.com — Entirely closed source; no published components, model cards, or security whitepapers

highVerified: 2026-07-09

community activity

Community engagement analysis of recipe ecosystem and growth signals

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Community members have built thousands of custom automations with a creator payout program ($0.10-$1.00 per recipe signup); user base reportedly 10x'ed in the months after public launch

mediumVerified: 2026-07-09

⚙️Operational Excellence

ease of integration

Onboarding friction assessment; messaging-native is the lowest-friction agent interface shipped to date

Evidence

9to5Mac - Apple's Messages app on iPhone now has a third-party AI agent — Zero-install onboarding: text Poke like a contact in Messages, Telegram, or WhatsApp; since 2026-06-04 it appears natively in Apple Messages for Business for iMessage's ~1B users

highVerified: 2026-07-09

scalability

Assessment of team size, infrastructure signals, and stated cost constraints against growth

Evidence

TechCrunch - Apple approves Poke as the first AI agent on its Messages for Business platform — 10-person team scaling to rapid consumer growth plus per-user Apple platform billing; real-time inference costs are acknowledged as the main constraint

lowVerified: 2026-07-09

cost predictability

Pricing model analysis; negotiated, usage-shaped pricing is affordable but unusually unpredictable

Evidence

TechCrunch - Poke makes using AI agents as easy as sending a text — Free for non-real-time use; paid pricing is flexible and usage-based, with beta users negotiating roughly $10-$30/mo (~$20 typical) via Poke's conversational pricing

mediumVerified: 2026-07-09

monitoring capabilities

Monitoring features assessment

Evidence

Poke — No dashboard, usage analytics, or audit tooling; the chat thread is the only window into what the agent has done

mediumVerified: 2026-07-09

production readiness

Vendor stability and maturity assessment: strong platform validation, very young company

Evidence

TechCrunch - Apple approves Poke as the first AI agent on its Messages for Business platform — Apple's first-ever AI-agent approval on Messages for Business validates the product, and $25M raised at a $300M valuation (Spark Capital, General Catalyst) funds it, but it remains a 10-person startup prioritizing growth over profitability

mediumVerified: 2026-07-09

Strengths

+Lowest-friction agent interface shipped: lives entirely in iMessage/SMS, Telegram, and WhatsApp with no app
+First third-party AI agent approved on Apple Messages for Business (2026-06-04), adding verified identity on iMessage
+Best-fit model routing across providers and open-source models instead of single-vendor lock-in
+Persistent memory and proactive trigger-based automations across 40+ integrations (Gmail, Calendar, Notion, Hue, Oura, GitHub)
+Thousands of community-built recipes with creator payouts
+Affordable: free tier plus flexible negotiated pricing around $10-$30/mo

Limitations

!Novel risk surface: standing access to email, calendar, and purchases through a channel with no OS-level permission model or revocation UI
!Email-triggered automations expose the agent to indirect prompt injection from any sender, with no published hardening
!Security claims (token isolation, pen testing) are vendor statements without independent audit
!No action log or review step; the chat thread is the only record of what the agent did
!Closed source, no compliance certifications, consumer-grade privacy policy
!10-person company: continuity, support, and incident-response capacity are unproven
!Negotiated usage-based pricing makes long-term costs hard to predict

Metadata

license: Proprietary

supported models

0: Multi-provider best-fit routing (major frontier providers and open-source models; internals undisclosed)

architecture: Cloud-hosted persistent agent reached via messaging channels (Linq-powered delivery); server-side automations with per-service OAuth connectors

deployment type: Cloud-only; interface via iMessage/SMS, Telegram, WhatsApp (EU/Brazil availability constrained), and Apple Messages for Business

tool support

0: Email (Gmail, Outlook)

1: Calendar

2: Smart home (Philips Hue, Sonos)

3: Health (Strava, Oura, Fitbit, Withings)

4: Productivity and dev tools (Notion, Linear, GitHub, PostHog)

first release: Beta late 2025; public launch March 2026; Apple Messages for Business approval 2026-06-04

pricing: Free tier; flexible negotiated/usage-based paid plans (~$10-$30/mo, ~$20 typical)

company: The Interaction Company of California, Palo Alto; founders Marvin von Hagen and Felix Schlegel; $15M seed (2024) + $10M (April 2026) at $300M post-money; backers include Spark Capital, General Catalyst, the Collison brothers, Guillermo Rauch

Use Case Ratings

research assistant

Handles quick lookups and standing news/price watches over text, but has no browsing transcript or citation depth

content creation

Drafts emails and short messages well; the texting interface is unsuited to long-form work

data analysis

No analysis environment; limited to summarizing data its integrations expose

education

Conversational tutoring and reminders over text suit casual learning; no structured curriculum tools

Similar Agents

Manus

Butterfly Effect / Meta (acquisition being unwound as of June 2026)

Zapier AI Actions

Zapier

Sierra

Sierra Technologies Inc.

MCP Gmail Server

Community

MCP Google Calendar Server

Community