Shortwave

The assistant writes into systems Shortwave does not own, through a named catalog. From the AI Integrations section of the assistant guide, each with its action: HubSpot, to access customer information and update CRM data; Notion, to read, create and update pages and databases; Slack, to access channels and send messages; Asana, to create and assign tasks; Linear, to track and update tasks; and Zapier, to connect to thousands of other services. Each is enabled with one click from a settings page, with authentication handled through Arcade.

Updating a CRM record, creating and assigning an Asana task and sending a Slack message are authenticated writes. The published examples cross systems: "Create Asana tasks for each of the bug reports in my inbox" and "Find all emails related to Project Falcon and create a status update in a new page in Notion". Gmail, Google Calendar and Google Tasks are the native base, with calendar invites created and events edited from the assistant, and the plans matrix adds auto BCC for logging mail into HubSpot and Salesforce.

Tasklet, which connects to more than 3,000 apps, is a separate product at tasklet.ai that Shortwave's pricing page describes as something Shortwave integrates with, not something a Shortwave seat includes.

Sourceshortwave.com/docs/guides/ai-assistant AI Integrations and shortwave.com/pricingread 2026-09-13

Multi step execution is documented on Shortwave's own surface. The assistant guide's worked example is a three stage chain issued as one instruction, "Find emails discussing the downtime last week, schedule a meeting with the people involved, and draft an FYI email about the meeting": retrieval, then a calendar write, then composition, each stage feeding the next.

Other examples chain retrieval into external writes, such as "Search my email for recent product updates and create a blog post outline" and "Create Asana tasks for each of the bug reports in my inbox". The launch post for the current assistant describes it making plans and taking multiple steps, running several AI powered searches and calendar lookups, backtracking to try different approaches and working around errors.

AI filters are the part that outlasts a request. They "automatically take action on new emails based on a custom AI prompt, including labeling, archiving, deleting, starring, and more", metered at 3, 10 and 50 by plan, and each is a standing rule that evaluates and then acts without a person present.

There is no workflow builder, no branching or conditional composition, no delegation between agents and no named runtime. Chains are assembled per request by the assistant, and the persistent unit is a single condition filter. Tasklet's autonomous automation belongs to a separate product and is not part of this.

Sourceshortwave.com/docs/guides/ai-assistant, shortwave.com/pricing AI filters row and shortwave.com/blog/new-shortwave-ai-email-assistantread 2026-09-13

Shortwave keeps a maintained search index of the mailbox, and three pieces of its own evidence show it. First, the security guide names Pinecone, a vector database, among only four sub-processors (Google Cloud, OpenAI, Anthropic and Pinecone), which discloses a stored embeddings layer by naming the vendor that holds it.

Second, the index is sold with a retention horizon that varies by plan: AI powered search history is 5 years on Business and unlimited on Premier and Max, described as finding emails with a simple description rather than exact keywords across the team's email history. Third, retrieval breadth is metered too: maximum threads per AI search is 50, 100 and 150 by plan, which only means something if a ranked index is being read.

The assistant docs describe an AI search infrastructure and an about:[topic] query that finds threads by description rather than keyword, and a dedicated Import full history how to exists for backfilling the corpus. Grounding is traceable: when the assistant answers from email it lists the relevant threads, with an about: button that opens the full set it drew on.

Sourceshortwave.com/docs/guides/security sub-processors, shortwave.com/pricing plans matrix and shortwave.com/docs/guides/ai-assistantread 2026-09-13

A real confirmation gate ships on one surface, and the surfaces that write to other people's systems have none. Asking the assistant to organize the inbox produces recommendations rather than actions: "It will recommend actions you can take quickly on groups of threads like archiving low-priority emails, adding labels, deleting junk, and creating todos."

The guide states the point: "This lets you get to inbox zero quickly while still staying fully in control at each step." The iOS release notes record the gate being added: "AI Assistant now waits for confirmation while organizing inbox." Drafts are always shown for review before sending and can be refined by follow up instruction, and the assistant lists the threads it drew on behind an about: button, so the user can check what an answer was built from.

The gate covers inbox organization only. The integrations create and assign Asana tasks, send Slack messages and update HubSpot records with no approval step documented. AI filters by definition "automatically take action on new emails", including deleting and archiving, unattended and without confirmation. There is no administrator layer: no approval queue, policy engine, risk tiers by action, or scoping of what an assistant may touch. Oversight is per user and per surface rather than governed.

Sourceshortwave.com/docs/guides/ai-assistant organize and integrations sections, shortwave.com/pricing AI filters row, and the vendor's App Store release notesread 2026-09-13

Shortwave has an attestation but no access controls a customer administrator can use. The vendor publishes CASA Tier 2 compliance, reviewed by third party security auditors and maintained by an annual security audit, and lists it as a line in the plans table. The security page's SOC 2 Type II and GDPR section says only that an organization requiring them should "contact our team", which asserts no certificate.

Customer access controls are missing. The pricing page's Privacy & Security section has exactly three rows: CASA Tier 2 compliance, block spy trackers and privacy mode. No SSO, SAML, SCIM, user provisioning, roles, groups, permissions model, admin console or customer visible audit log appears in the matrix or the security guide, and there is no enterprise tier for such controls to sit in.

The security page's Audit logging section governs Shortwave's own employees: access to customer data is tightly controlled internally, all such access is audit logged, no employee accesses data without explicit customer permission, and employee access requires hardware security keys.

Those are strong internal controls, not a surface the buyer's administrator operates. Google Advanced Protection and the Google Workspace Marketplace listing are Google's programs, enforced by the buyer's own Workspace administrator, and thread sharing, comments, assignees, shared labels and channels are collaboration, not access governance.

Sourceshortwave.com/pricing plans matrix and shortwave.com/docs/guides/securityread 2026-09-13

Nothing published lets a buyer reconstruct what the agent did. No run history, activity log, admin dashboard, export, retention statement or customer visible audit trail appears in the docs or in the plans matrix.

The security guide's Audit logging heading is about Shortwave's own employees: "Access to customer data is very tightly controlled internally and all access is audit logged to ensure compliance", with no employee access without explicit customer permission and hardware security keys required. That is the vendor auditing its staff; nobody outside Shortwave reads that log.

Three features look like observability and are not. The assistant listing which threads it searched, behind an about: button, lets a user check one answer, not the agent's conduct. Read statuses, recent opens and link tracking tell the user when a recipient opened an email or clicked a link, which measures the recipient. Shared labels are a team filing structure. Most important, there is no customer facing record of assistant runs, filter executions or integration writes, including what AI filters deleted and archived unattended.

Sourceshortwave.com/docs/guides/security audit logging section and shortwave.com/pricing plans matrixread 2026-09-13

AI Memories let a user steer the assistant by telling it what to remember. In the vendor's framing, memories "control the AI assistant's behavior, including how it responds, and what actions it takes along the way", and the plans matrix lists the feature on every paid plan as "Customize the behavior of your AI assistant by asking it to remember facts & preferences". Memories carry across sessions and the assistant consults them as it works.

The user writes the entries in plain language, and every published example is a standing instruction a person typed, from "Remember to use Oxford commas" to "Remember to add Asana tasks to the Marketing project". Nothing the agent learns during a run is written back. No page shows where memories are kept or how a user would view, edit or delete one, and no retention, expiry or purge path is published for a memory apart from deleting the account.

Ghostwriter learns a writing voice from sent mail, and context carried from earlier searches into follow ups stays within one conversation.

SourceShortwave, shortwave.com/docs/guides/ai-assistant AI Memories section, shortwave.com/pricing and shortwave.com/changelogread 2026-10-05

Shortwave names its host but no place, and offers no choice. The one published sentence is "All Shortwave data is stored in Google Cloud's highly secure & compliant data centers", with AES256 at rest and TLS 1.2+ in transit. There is no region, region list, customer environment option or selection surface, and no residency is claimed at all: not a country, a continent or an EU option.

Shortwave is delivered as cloud SaaS with native apps on iOS, Android, Mac and Windows; where the buyer installs the reader says nothing about where the service runs. No on premises, private or dedicated tenancy, bring your own cloud or air gapped mode exists, and there is no enterprise tier in which to sell one; the three published plans are Business, Premier and Max.

The absence is structural. The product is a Google Workspace Marketplace app that syncs a customer's Gmail into Shortwave's own Google Cloud estate and indexes it there, with OpenAI, Anthropic and Pinecone as named sub-processors.

Sourceshortwave.com/docs/guides/security data storage and sub-processors sections, shortwave.com/pricingread 2026-09-13

Shortwave ships nothing prebuilt; every reusable unit is written by the customer. Saved prompts are built by the user: "Set them up in the Assistant settings by choosing an emoji, naming your action, and defining your prompt." The single example, a coffee chat scheduler, is illustrative text in the documentation rather than an entry in a library.

Snippets let a user "create smart templates for writing common email content" and share them with a team, so the team library is whatever that team wrote. AI filters are natural language rules the customer writes, metered by count, and AI memories are instructions the customer writes.

There is no gallery, template library, starter pack, agent directory or marketplace of any kind on the site. Sharing infrastructure for content users create is not a shipped catalog. The three plans are a pricing structure, and the named integrations are connectors rather than agents.

Sourceshortwave.com/docs/guides/ai-assistant saved prompts section and shortwave.com/pricing snippets and saved prompts rowsread 2026-09-13

One event starts an agent, mail arriving, and nothing else does. AI filters, per the plans matrix, "automatically take action on new emails based on a custom AI prompt, including labeling, archiving, deleting, starring, and more", metered at 3, 10 and 50 by plan. That is an unattended run started by an inbound event with the model in the loop, not a static rule. Auto apply labels and delivery schedules sit alongside it, and instant AI summaries generate automatically on every thread.

The assistant is reachable on web, desktop, an installable PWA, iOS and Android, by Cmd-J or the sparkle icon on each, with granular push notifications. Those are places to invoke it, not triggers.

No scheduled or cron runs, webhook, inbound API, external work queue or monitoring mode that starts an agent is published in the docs or the plans matrix. Tasklet, which runs background automations around the clock, is a separate product at tasklet.ai that Shortwave's pricing page describes as something it integrates with; its scheduler is not a Shortwave trigger.

Sourceshortwave.com/pricing AI filters and delivery schedules rows, shortwave.com/docs/guides/ai-assistantread 2026-09-13

The user chooses the model. The pricing page says so in its own words: "AI Intelligence: choose between Basic (Claude Haiku 4.5), Standard (Claude Sonnet 4.6), Advanced (Claude Sonnet 4.6 with adaptive thinking), and Expert (Claude Opus 4.6 with adaptive thinking) intelligence levels." Four named models are exposed as a setting. The quota row confirms the choice is per request: "Quota usage varies based on the complexity of the request & the intelligence level selected."

Plans set the ceiling: Business tops out at Standard, Premier at Advanced and Max at Expert, and the choice beneath it is the customer's.

Every model on the selector is from Anthropic. OpenAI appears only on the sub-processor list, which does not make it a second selectable model. The vendor also states that most of its AI workloads run open source models on hardware it controls; those are not named or selectable. So this is a single provider selector with four intelligence levels.

Sourceshortwave.com/pricing AI Intelligence and quota rows, with shortwave.com/docs/guides/security sub-processors as corroborationread 2026-09-13

A customer can extend what Shortwave reaches, but nothing lets an outside system drive Shortwave. The assistant guide publishes "Connect to any tool with custom MCP servers": point to an MCP server of your choice, "or even create your own". A server the customer writes and Shortwave then calls is a real extension surface.

Shortwave is the MCP client, not the server. The how to tells the user to provide the command, arguments and optional environment variables for a server, to check ~/Library/Logs/Shortwave/main.log on Mac or %APPDATA%\Shortwave\logs\main.log on Windows when it fails, and to make sure they have the latest desktop app: a local process started by Shortwave's own client. It also publishes the client's limits: MCP support is limited to tool calls, resources, prompts and sampling are unsupported, and only a limited number of tools (currently 40) can be active at once.

The documentation has four sections (Guides, Migrations, References and How-tos) and no developer portal, REST or GraphQL reference, SDK, webhook, CLI or A2A card. There is no docs or developer subdomain. The nearest thing to a public interface is the Google Workspace Marketplace listing, which is a distribution channel rather than an API.

Sourceshortwave.com/docs/how-tos/using-mcp and shortwave.com/docs/guides/ai-assistant custom MCP sectionread 2026-09-13

No way for a customer to evaluate a change is published. There is no test harness, scored cases, dry run, sandbox, regression check, A/B comparison of configurations or published quality metric.

That matters more here than usual. AI filters are rules the customer writes in plain English that "automatically take action on new emails, including labeling, archiving, deleting, starring", and the plans sell 3, 10 and 50 of them. The first time a filter runs is against live mail, with deletion among the available actions. Nothing documented lets a customer test that prompt against past threads, preview what it would have caught, or see what it did afterward.

The daily AI usage quota is consumption metering, not evaluation. Intelligence levels let a user pay for a better model, which is model choice, not a comparison. Refining an answer with a follow up instruction is iteration within one conversation, not a test of a change.

Sourceshortwave.com/pricing plans matrix and shortwave.com/docs/guides/ai-assistantread 2026-09-13

The web browsing feature gathers information; it does not operate websites. The assistant guide lists the browser as one of seven one click AI integrations and states its job: "Web Browser: crawl the web to explore links and gather information." The plans matrix repeats it as AI web browsing that lets the AI "browse the web to get the latest information related to your emails".

No hosted or local browser session the agent drives, desktop control, or remote or local computer control is documented. Shortwave reaches every system it touches through APIs: the Gmail and Google Calendar APIs as a verified Workspace Marketplace app, OAuth connectors to HubSpot, Notion, Slack, Asana and Linear brokered through Arcade, Zapier, and MCP tool calls for anything else. The desktop and mobile apps are Shortwave's own clients, not the agent operating someone else's interface, and the MCP host starting a local server process is a runtime detail.

Sourceshortwave.com/docs/guides/ai-assistant AI Integrations and shortwave.com/pricing AI web browsing rowread 2026-09-13