Cassidy vs Lindy (2026)

Breadth specified rather than asserted as a round number. More than a hundred pre-built workflow actions cover content generation, Knowledge Base search, email, Salesforce and HubSpot record updates, document creation, Slack messaging, conditional routing and looping, and the vendor names Google Workspace, Microsoft 365, Slack, Salesforce, Notion, SharePoint, HubSpot, Airtable, Zendesk, ServiceNow, Jira and Zapier among connected tools. That spans the classes properly: CRM, ticketing, ITSM, messaging, productivity, storage and project management. What makes it more than a connector count is that integrations work in both directions. The same systems act as knowledge sources feeding the Knowledge Base, as trigger sources firing workflows on a new Zendesk ticket or a Salesforce record change, and as action targets the agent writes to. A single Salesforce connection is context, trigger and destination at once, which is a tidier model than platforms where each direction is configured separately. Webhooks, an API and custom connectors over MCP cover anything unlisted, so tools that expose themselves only over MCP are reachable too.

Connections span Gmail, Slack, Notion, HubSpot and 1,000+ tools plus any hosted MCP server, with a credentials vault, per-integration guardrails, an HTTP Request skill for anything uncatalogued and 500+ actions from a Pipedream partnership, though legacy material claims 3,000+ integrations.

The control-flow primitives are named rather than asserted. The action library carries dedicated categories for AI, Data, Knowledge Base, Control Flow and Human in the Loop, with Paths for conditional branching, Loop for iteration and Combine Text for merging, plus documented error handling that retries failed steps automatically or routes errors for review. Two capabilities lift this above a competent workflow builder. Workflows can run across large datasets, with the vendor's own example being a thousand documents processed at once, which is the batch dimension most no-code builders omit and the one that decides whether a workflow survives real volume. And version control is documented: changes are tracked, iterations managed, and a workflow can be rolled back, with the Copilot builder offering a restore-checkpoint action. The distinction the vendor draws in its own action list is a good sign of design thinking, separating Generate Text, which runs without business context, from Generate Text with Cassidy Assistant, which carries the Knowledge Base. A builder can choose per step whether reasoning is grounded, rather than the platform deciding for them. Agents complete tasks and take action across tools, updating Salesforce records, posting to Slack and routing tickets, so orchestration ends in action rather than output.

One agent plans and runs multi-step work across tools, with Routines and skills, but no branching, conditions, multiple agents or flow surface the customer configures is documented.

All three trigger classes are documented. Voice is not the bar; the bar is coverage across events, schedules and channels, and Cassidy has all three. Schedules and events are explicit. Workflows run on a recurring schedule, fire from a webhook, accept mail at a dedicated inbound email address, start manually from a form, and fire on integration events including a new Zendesk ticket, a HubSpot or Salesforce record change, a Jira issue, or a meeting ending. That last one is unusual and worth carrying: a meeting concluding is a business event most platforms cannot observe at all, and Cassidy can because it records the meeting. Channel coverage is broad: Slack, Microsoft Teams, Google Chat, email, browser extensions for Chrome and Edge, Microsoft Word, Excel and Outlook, and a website embed. Agents reach people where work already happens rather than requiring a visit to Cassidy. So this is a strong Full rather than a marginal one: work reaches the agent by time, by event and by conversation.

Schedules through Routines, events such as a meeting ending or an unanswered email, and channels across Slack, iMessage and SMS, Gmail and the major meeting platforms are all included on every plan.

It clears the persistence line by a distance and adds two unusual things. Persistence is unambiguous: the Knowledge Base syncs from Google Drive, SharePoint, Confluence and Box and is kept current automatically as files are added and edited, so it is a standing structure that tracks its sources rather than context assembled per prompt. Semantic search runs over it, and sources can be grouped into organized collections. The first distinguishing feature is permission awareness: the vendor states users can only view documents and receive answers from sources they are authorized to access. That closes the privilege-escalation path that makes enterprise search dangerous. The second is Document Verification. Teammates can verify a source, request that someone else verify it, or flag it as outdated, so the corpus has a human-maintained freshness signal rather than decaying silently. Combined with every response citing the exact source it used, a reader can check a claim and a subject-matter expert can act on what they find. Grounding here is not just retrieval, it is retrieval with an accountability loop around it.

Recorded meetings filed into a searchable library of shareable folders give a maintained knowledge store, while answers across Slack history and connected tools come back with the underlying file or decision attached.

Conversation memory on the Generate Text and Generate Text with Agent actions lets the AI remember previous messages across Workflow runs, in a single shared thread or in keyed threads grouped by a variable such as a ticket ID, email thread or customer. The AI's response is saved to the history by default, a read-only option exists, and a keyed thread can be shared across Workflows. That is agent-written state persisting across runs with a documented scope, which is Partial. What would take it to Full is a documented lifetime or purge path and a way to review or delete the memory; data retention settings are documented for run history and agent chat history, not for this memory.

Memory is a versioned plain-file system the customer can open, read and edit like a document, scoped personal or workspace, and skills keep demonstrated procedures for reuse across the team.

The mechanism is textbook. Approval steps can be inserted anywhere in a Workflow, and the vendor's own description is exactly what this axis asks for: they pause execution until a person reviews the output, provides input, or approves the next action, intended for quality checks, sensitive decisions, or anywhere human judgment is required before proceeding. The action library carries a dedicated Human in the Loop category alongside AI, Data, Knowledge Base and Control Flow, so this is a first-class construct rather than a pattern a builder improvises. A runtime guardrail enforcement engine is not the bar. The published bar asks for the shipped approval mechanism, and a pause-for-sign-off step placed at the builder's chosen point is the most direct form of it. The supporting layer is real too. Document Verification lets teammates verify, request verification of, or flag as outdated the sources an agent cites, which is oversight of the grounding rather than the action. Permission-aware retrieval means an agent cannot surface what its user is not entitled to see, and granular role-based access governs who may build and deploy at all.

Anything with outside impact, such as sending an email or updating a ticket, waits for a named approver on every plan while read-only lookups from approved sources proceed, with per-integration guardrails and permissions the customer sets and revokes.

Both halves of the bar are met plainly. Cassidy states it is SOC 2 Type II, GDPR, HIPAA and CASA certified, with no hedging: no aligned with, no in progress, no supports compliance. Plain assertion, rather than alignment language, is what the bar asks for. The control surface is coherent and enterprise-appropriate: encryption in transit and at rest, granular role-based access control, organization and workspace level permissioning, single sign-on, and full audit logging. Permission-aware retrieval is the control that matters most for what this product does, because it means the Knowledge Base cannot become a way for a user to read documents they could not open directly. The posture is stated twice and unambiguously: customer data is never used to train models, and Knowledge Base content is used solely to provide context for that customer's Agents and Workflows. For a product whose entire value rests on ingesting a company's internal documents, that is the first question a buyer asks, and answering it in plain language on the knowledge-base page rather than burying it in a policy document is the right instinct.

SOC 2 Type II audited by Johanson Group with a Drata trust center that releases reports on request, per-integration permissions and encryption on every plan, and SSO and SCIM on Enterprise only.

A per-step record of what ran. Progress appears under Steps as each action finishes, a completed run expands to show every action, and View Previous Runs lists past runs with the credits each consumed, including runs fired by schedules and integrations. The pricing page lists an audit trail of runs as a full log of Workflow executions, Browser Use records the full sequence of actions with a screenshot per step, and organization audit logs record who viewed or changed Agents and Workflows, with search, filter and export. Retention of run history is configurable.

Team files carry version history and the meeting library is searchable, but no logging, tracing or audit page is documented for the current product, and a per-run task view appears only in legacy builder documentation.

Cassidy is a multi-tenant cloud platform sold self-serve with a credit model, and nothing on any surface reached documents self-hosting, on-premises installation, virtual private cloud deployment, customer-managed encryption keys or a customer-selectable region. What exists instead is logical rather than physical separation: organization and workspace level permissioning, granular role-based access, and permission-aware retrieval. Those govern who sees what inside Cassidy's environment, which is a security property and counts there. They do not answer where data lives, which is what this axis asks. The absence is coherent with the product and worth stating that way rather than as a failing. Cassidy is certified SOC 2 Type II, GDPR, HIPAA and CASA and encrypts in transit and at rest, which is the route a self-serve SaaS product takes to satisfy regulated buyers instead of offering single-tenant deployment. An enterprise tier could offer dedicated hosting without it appearing on the enterprise page, though nothing in the pricing structure suggests it.

The security page says residency requirements are supported, but no region, selection control, VPC, on-premises, self-hosted or dedicated option is documented, so the product is managed SaaS only.

A browsable catalog is evidence of the bar rather than its definition: what qualifies is packaged assets the vendor ships ready made, adoptable by any route, and a templates page is one of those routes. Cassidy ships more than eighty agent templates, which is a named set of ready-made agents a customer starts from rather than components they assemble. Alongside them sit reusable prompt templates stored and reused across the workspace, and more than a hundred pre-built workflow actions covering content generation, knowledge search, email, CRM updates, document creation, Slack posting, conditional routing and looping. A Copilot builder generates a working trigger and action chain from a plain-language description, which is a different route to the same outcome: the customer receives something working rather than a blank canvas. The Workflow templates catalog is browsable by category and searchable by keyword, with guided setup that authorizes integrations and selects the Agent. The full catalog was not opened, so the eighty-plus count rests on the vendor's own claim.

The product ships 40+ skills on every plan that the agent picks up on its own, and customers add their own by showing Lindy a task once, after which it is saved for the whole team.

The specifics are strong. The vendor documents more than thirty models from OpenAI, Anthropic and Google, selectable per action, with smart recommendations indicating which suits each action type so a non-technical builder is guided rather than left guessing. Selection per action rather than per agent is the finer granularity and the more useful one: a single workflow can use a cheap model for routing and an expensive one for the final draft. Models are organized into Standard and Premium tiers, which is worth carrying to a comparison page because the vendor is explicit that the tier chosen is the main driver of cost under the credit model. That makes model choice a budget control rather than only a quality one, and it is unusually transparent about the trade-off. The no-lock-in commitment is documented and specific: as new models are released, existing Workflows gain access and an immediate upgrade, with no separate provider API keys required. For a no-code product aimed at non-technical teams, removing key management is what makes multi-model real rather than nominal, since the alternative is a business user being asked to obtain credentials from three providers.

The customer picks a model per task from a multi-provider roster, with model selection listed as included on every plan and the changelog recording models added over time, such as Gemini 2.5 Flash-Lite for phone calls.

A documented API is the bar on this axis, and Cassidy publishes one: organization API keys created and deleted in Organization Settings, a thread-and-message model to create conversations and send messages, sample code on each agent's Deploy via API tab, and an endpoint reference at app.cassidyai.com/api-docs. Inbound webhook triggers make a Workflow startable by any external system, which is being callable from outside in the most practical sense, and Workflows can be imported as JSON. Agents also embed on a website and extend into Chrome, Edge, Word, Excel and Outlook, so the platform reaches into surfaces the vendor does not own. No SDK package or first-party MCP server is published; MCP is corroboration on this axis rather than the requirement, so its absence does not withhold Full.

The webhook trigger takes bearer-token authenticated calls with a callback response and an HTTP Request skill calls outside APIs, but no SDK, client library or Lindy-published MCP server is documented, and the OpenAPI specification listed in the documentation index could not be read.

A testing and debugging surface rather than an evaluation harness. A Test Workflow run uses sample inputs or a real recent record pulled from a connected app and shows results for each action; a single step can be retested, or the run tested from a step or up to a step; nested actions inside Paths and Loops can be tested; and actions such as sending email can be skipped during tests. Skills can be tested on difficult cases before sharing and compared across versions, and version history supports rollback. What Full would need is a scored test set, judge, metric or comparison of outputs between versions on the same inputs, and none is documented.

Neither the documentation index nor the pricing page's included capabilities list any test, preview, sandbox, evaluation or benchmarking surface, and routines and skills take effect directly with no documented dry run.

A platform-shipped browser the agent operates. Browser Use lets an Agent automate interactive web tasks in a live browser, navigating pages, clicking buttons, filling out forms and extracting data, for sites without APIs such as government portals and supplier platforms, and Workflows can call it too. The session runs live with real-time status and a side-by-side preview, the user can take control to solve a CAPTCHA, and the finished run shows the full sequence of actions with a screenshot per step. The vendor distinguishes it from Web Search, which only reads pages.

Computer use is included on every plan: the vendor introduced it as Autopilot, giving the agent its own cloud computer to do what the user could when a tool is not otherwise reachable.