Carly

The counterparty estate is deep and named by product. The integrations page lists AppFolio, BoldTrail, Buildium, CINC, DoorLoop, Entrata, Follow Up Boss, Guesty, Hospitable, Hostaway, Lofty, Propertyware, Real Geeks, RealPage, Rent Manager, Sierra Interactive, Wise Agent and Yardi in property alone, with Bullhorn, CompanyCam, Deel, Greenhouse, JobTread, Outreach, Procore, Ramp, Salesloft, Smartsheet and Teamwork alongside, and CRM coverage across HubSpot, Salesforce, Attio, Dynamics 365, Zoho and Pipedrive.

The estate is unbounded by design: "Carly reaches these with your own API key instead of a built-in connector. Paste your key on the integrations page in your Carly dashboard and they work the same way everything above does." That is the customer's key for a third party service, an integration mechanism, not an API for Carly.

The direction is write, not read. The MCP tool table has agents "send and reply from your own mailbox," "create, update, move and delete events," "add and update contacts," "create and append to docs," "append workbook rows" and "create folders and share" across Google Drive and OneDrive. Writing a calendar event and appending a workbook row are authenticated writes into systems of record.

Each agent receives its own configured tool access rather than inheriting a global connection set, so the integration estate is partitioned by actor.

The integration count varies across Carly's own pages: 120+ in the homepage meta description, 200+ in one blog post and "260+ across 45+ categories" in another. The named estate and the own key path say more than any of those numbers.

Sourceusecarly.com/integrations/ named counterparties and own-key mechanism, usecarly.com/mcp/ tool tableread 2026-09-13

The customer composable workflow layer is documented first party, with its full lifecycle: "author workflow graphs or start from a template, list, test, run, pause, resume, archive, schedule recurring tasks." A graph the customer authors, tests, runs, pauses, resumes and archives is orchestration in the strict sense.

Control flow is explicit. The vendor states "Carly now does what Zapier does: triggers, actions, branches, 200+ integrations," and branches are conditional execution. A /workflows/ section carries an overview, a template gallery and thirteen categories, from inbox and calendar through sales, support and internal ops to simple (no AI).

The multi actor case is the agent model itself. A customer creates named agents from a dashboard, each with its own instructions, configured tool access, permissions and email address, so work is divided between specialized actors that can be copied onto the same thread.

Execution is multi step and crosses systems. The published appointment setting behavior runs end to end without the user: a reschedule request reopens the negotiation, finds a new slot, updates the calendar and confirms, and a missed meeting produces a same day follow up offering to rebook. The homepage instructions are compound by design, such as "enrich new form leads and email me" and "pull the earnings reports and save them to a folder," each a chain of retrieve, transform and write across different systems.

Sourceusecarly.com/mcp/ Workflows tool group, usecarly.com/workflows/ template categories, usecarly.com agent modelread 2026-09-13

Agents can search real corpora, but no retrieval layer, index or citation behavior is documented as a product surface.

The searchable estate is broad and enumerated in the MCP tool table. Mail: "search and read mail." Contacts: "look up people, search, query and aggregate CRM records," where aggregate is computation over a body of records rather than lookup. Structured data: "create data tables and query rows." Notes: save, search and delete.

Files: Google Drive and OneDrive list and download, so documents in the customer's storage are reachable. Each agent also carries its own files. The built in lightweight CRM populates contacts automatically from email and calendar activity, with natural language search over that history, a corpus the platform maintains rather than one the customer curates. Contact enrichment and web research extend it outside the tenant.

Nothing describes an index, embeddings, vector store, semantic ranking, chunking or a knowledge base product, and nothing describes citation. No page says an agent's answer points back to the message, record or file it came from. On a product whose agents write to clients on the customer's behalf, an unsourced assertion in an outbound email is the failure that matters, and there is no published mechanism against it. Search over a database is not a retrieval layer.

Memory save and search is scoped per agent, and the connector estate is what fills these stores.

Sourceusecarly.com/mcp/ tool table, usecarly.com CRM descriptionread 2026-09-13

Carly has an approval gate on one path only.

On the email path there is no approval step by design, and the vendor sells its absence: "many AI tools draft the reply and wait for you to hit send, whereas Carly sends and follows up on its own."

On the MCP path there is a real mechanism: "Read-only tools run without interrupting you; tools that write ask first." That is a shipped confirmation gate with a documented scope rule, interposing a person on writes. It works per action rather than per change, so a person reviews and approves each write.

A second layer bounds agents before execution. Each agent carries its own configured tool access and permissions, so what an agent may do is decided by a person before it does anything, and OAuth scoping binds every call to the authorizing account.

That gate covers writes initiated through an MCP client, not the product's headline behavior. An agent that decides on its own to send a follow up to a client, from the customer's domain, has no documented approval queue, confirmation step, per action risk tiering or configurable autonomy level. Nothing lets a buyer require sign off before any outbound mail. Oversight exists where a person is already driving a client, and is absent on the autonomous path where it would matter most.

Sourceusecarly.com/mcp/ authentication and tool-behavior statement, usecarly.com agent permission modelread 2026-09-13

Access controls are documented, but no attestation is published.

No SOC 2, ISO 27001, GDPR, HIPAA or CCPA claim, auditor, report, penetration test, trust center or compliance or security page appears anywhere on an estate with a seven menu header, a five column footer, and legal reduced to Terms and Privacy on calbotservice.com. The one page that reads like security, /solutions/it-security/, is a use case page about serving IT and security teams, not a statement of Carly's own posture.

Access runs on "OAuth 2.1 with PKCE. No API key to paste and no shared credential. Every connection is bound to one person's Carly account... every call is scoped to the account that authorized it." Delegated authorization with per user binding and explicit rejection of shared secrets is a real customer facing control. Each agent also carries its own configured tool access and permissions, so authority is scoped per agent rather than granted globally.

An agent holding OAuth grants to a user's mailbox, calendar, CRM and cloud storage, sending from the customer's own domain, is a broad trust grant, and this vendor publishes no attestation at all. No SSO, SCIM or RBAC is documented either.

Sourceusecarly.com/mcp/ authentication section, usecarly.com full navigation and footerread 2026-09-13

Because each agent has its own email address and works in the thread, its actions are visible in the inbox by construction to everyone on that thread. That is observability arriving as a property of the interaction model rather than as a feature, and it is unusually legible: the artifact a buyer inspects is the actual email the agent sent, in the same place a colleague's would be. External parties see the agent's work too, which no dashboard achieves.

A second surface is documented: workflows can be listed, paused, resumed and archived through the MCP tool table, so a customer can enumerate what is configured and running and halt it, programmatically as well as in the dashboard.

No run history, execution log, decision or reasoning trace, error or failure surface, retry visibility or audit log of agent actions is documented. The inbox shows what the agent sent; it does not show what the agent considered, what it read to decide, which tools it called, or what it tried and failed.

That gap matters more than it first appears. This agent sends mail autonomously from the customer's domain, so the actions least visible in the inbox (a CRM record silently updated, a file moved, a tool call that errored) are the ones with no trace. The lightweight CRM's contact history describes the customer's business, not the agent.

Sourceusecarly.com agent email-address model, usecarly.com/mcp/ workflow list, pause, resume and archive toolsread 2026-09-13

Carly's memory is a documented, operable object. The vendor states each agent gets its own memory, and the MCP tool table exposes it: "Notes, to-dos & data: save, search and delete memory, manage to-dos, create data tables and query rows." Save, search and delete is a read and write surface over retained state, exposed to an external caller and named as memory by the vendor.

Memory is per agent, and OAuth binds every connection to one person's account, so state is scoped rather than pooled. Delete memory is documented as a tool, which gives a purge path. What is missing is a published lifetime or retention policy, so nothing states how long memory persists, whether it expires, or what happens to it on cancellation, and no trust or data handling page exists to carry that.

Cross session continuity is independently evidenced. The vendor describes keeping an agent running after the user closes their laptop, and the behavior requires it: a missed meeting produces a same day follow up, and a reschedule request reopens a negotiation days later, which only works if the agent carries the thread's state forward.

The lightweight CRM is application data rather than agent memory.

Sourceusecarly.com/mcp/ tool table, usecarly.com agent modelread 2026-09-13

Nothing names a place or offers a choice. No region, region list, country, cloud provider, tenancy model, self hosted or on premises option, customer cloud path or selection surface appears on the homepage, the MCP page, the integrations page or in a footer with five link columns. The product runs as cloud SaaS and claims no residency at all.

Low deployment friction is not a deployment option. Carly is SaaS with no application to install, running over existing email and messaging channels; that speaks to onboarding effort, and having no application to install means the buyer receives nothing to place anywhere.

The estate spans three domains and none carries a hosting statement: usecarly.com for marketing, carlyassistant.com for the application and the MCP endpoint, and calbotservice.com for terms and privacy. The absence of any residency page is notable on a product whose agents hold OAuth grants to a user's mailbox, calendar, CRM and cloud file storage; an EU buyer has nothing to evaluate.

Sourceusecarly.com navigation and footer, usecarly.com/mcp/read 2026-09-13

The browsable catalog is a navigation section with its own URLs. usecarly.com/agents/templates/ is Agent Templates, carried in the header under an Agents menu and again in the footer, beside a Prompt Library at /prompts/ and a Use Cases index.

Separately, usecarly.com/workflows/templates/ is a workflow template gallery filterable across thirteen named categories: inbox and email, executive assistant, calendar and meetings, personal assistant, sales and pipeline, customer and support, marketing and content, internal ops, daily newsletters, data briefings, booking pages, fun and team, and simple (no AI). Filterable categories with durable query URLs make a catalog.

The selected unit does work when selected. The MCP tool table confirms the mechanic: "author workflow graphs or start from a template." Start from a template is the adoption action, and what arrives is a configured graph that runs.

Each unit stands on its own. Remove the daily newsletter templates and the sales and pipeline templates are still whole, unrelated working units; remove the client intake agent and the follow up agent still does its job. These are separable products doing different work, not tiers or modules of one thing.

Carly's tiers (Free, Personal and Enterprise) are a pricing structure rather than separate products. The free booking pages and the standalone utilities in the Free Tools menu are lead generation tools rather than agents.

Sourceusecarly.com/agents/templates/, usecarly.com/workflows/templates/ category filters, usecarly.com/mcp/ template toolingread 2026-09-13

Three invocation classes are documented.

Inbound message as trigger, with the agent addressable by anyone who can send mail. Every agent has its own email address, and the interaction is "just CC Carly, she'll take care of the rest." A third party on a thread, outside the customer's organization, can start agent work by replying. That is a genuinely external event surface, and it is why the vendor needs no separate app. Agents are also reachable via Telegram, WhatsApp, Slack and text.

Schedule, in two forms. Recurring instructions are the homepage's own examples: "send a reminder every Monday to check in with ops" and "check my emails every hour and let me know if anything's important." The MCP tool table confirms "schedule recurring tasks" as a first class operation, and scheduled check ins on dormant contacts are a published behavior.

Data condition and state change detection. A missed meeting produces a same day follow up offering to rebook, a reschedule request reopens the negotiation automatically, and new form leads are enriched on arrival. These fire on something happening in the world rather than on a clock, and Carly has both modes.

Unattended operation is the premise: "many AI tools draft the reply and wait for you to hit send, whereas Carly sends and follows up on its own," and the follow up on a missed meeting happens the same day whether or not the user opens anything. Through the MCP server, an external caller can also drive the platform.

Sourceusecarly.com homepage instruction examples and CC model, usecarly.com/mcp/ scheduling toolsread 2026-09-13

Carly discloses nothing about the model powering its own agents. No model, provider or family is named on the homepage, the MCP page or the integrations page, and there is no model selector, admin entitlement, per agent model setting, disclosed routing, or bring your own model or key path for inference.

The estate is full of model names pointing the other way: "Connect Carly to Claude, ChatGPT, or any MCP client," with setup blocks for Claude, Claude Code, ChatGPT, Cursor and VS Code, and a tutorials category headed "ChatGPT & Claude (MCP)." Each of those is the customer's assistant calling Carly through the MCP server, which is the customer choosing which AI tool drives Carly, not which model Carly reasons with.

The one adjacent fact is commercial: pricing is metered against an LLM credit allowance, and the free tier contains no AI, so runs are uncapped there because nothing consumes credits. That says inference is metered and still names no provider.

The customer's own API key also appears, but "paste your key on the integrations page" covers third party SaaS without a built in connector, not a model endpoint.

A vendor that names a single provider makes the absence of choice visible; Carly names none for its own stack while naming five for its clients.

Sourceusecarly.com/mcp/, usecarly.com homepage, usecarly.com/integrations/read 2026-09-13

Carly's own MCP server does the work of an API or SDK: its endpoint, its auth scheme and its tools are published first party, and the tools create or change the platform's own core objects. All four are on usecarly.com/mcp/, which sits in the main navigation and every footer. The endpoint is documented (https://carlyassistant.com/mcp/). Auth is OAuth 2.1 with PKCE, and every connection is bound to one person's Carly account with no shared credential.

122 tools are enumerated by category across email, calendar, CRM, workflows, files and notes. The Workflows group authors workflow graphs or starts from a template, and lists, tests, runs, pauses, resumes, archives and schedules them, so an outside caller composes and operates the platform's core object. The direction is inward: the tools drive Carly itself rather than bridging to the customer's other systems. No REST or GraphQL reference, OpenAPI spec or SDK is published on usecarly.com.

Sourceusecarly.com/mcpread 2026-09-23

Testing is a documented lifecycle verb, with no readable result. The MCP tool table lists the Workflows group as "author workflow graphs or start from a template, list, test, run, pause, resume, archive, schedule recurring tasks," so test sits beside run as a distinct operation, and a workflow can be exercised before it is committed to a schedule or a channel. The vendor also describes testing an agent in the browser before adding channels. Together they give a documented path for trying a configuration before it acts on live mail.

A test produces no readable, comparable result. There are no test cases, scores, pass or fail verdict, judge, golden set, regression suite, accuracy or quality metric, and no comparison between two versions of an agent. Nothing tells a buyer whether the agent got it right, only that it ran.

The gap matters more here than on most products. Carly sends email autonomously from the customer's own domain to their clients, and its differentiator is that it "sends and follows up on its own" rather than drafting. A platform that acts irreversibly on a buyer's behalf and publishes no way to measure whether it acts well has a real hole.

Credit and workflow run accounting is consumption reporting, and the write confirmation on MCP tools gates each action rather than a change.

Sourceusecarly.com/mcp/ Workflows tool groupread 2026-09-13

None of the three modalities is documented: no hosted or local browser session the agent drives, no desktop control, and no remote or local computer control. The architecture rules the category out.

Every route the agent takes into another system is programmatic: the MCP tool table lists 122 API backed tools across email, calendar, CRM, workflows, files, notes and booking pages, and the integrations page reaches anything else with the customer's own API key. The vendor positions API integration explicitly as the alternative to browser automation for recurring work. Google could redesign its interface entirely and a Carly agent would be unaffected.

Three things come close without being computer use. A Chrome extension is carried in the navigation and footer, but it is a browser tool a person installs and uses, not the agent operating a browser. The MCP tool table includes "run sandboxed Python and pandas," and code execution touches no interface at all. And an agent sending and reading mail from its own address is using a mail protocol, not driving a mail client's screen.

The absence fits the product, whose premise is "no app needed": the agent lives in the protocol layer and reaches people through email rather than operating anything visual.

Sourceusecarly.com/mcp/ tool table, usecarly.com/integrations/, usecarly.com no-app-needed positioningread 2026-09-13