Airtable vs Retool (2026)

Airtable has native integrations for Slack, Google Drive, Salesforce, Jira and Zendesk, with a full integrations directory and a public marketplace. Agents call document analysis, web search and image generation tools.

Sourceairtable.com/integrationsread 2026-09-01

Agents act through core tools built by Retool (email, calendar, Google Docs and Slides, files, web search), custom tools, any saved query or workflow, other agents and connected MCP servers, over Retool's 100+ connectors. There are 38 core tools in all, covering calendar events and open time slots, files and folders, vector search, reading a web page and running JavaScript or Python. An agent can connect up to 20 remote MCP servers and import tools from other agents.

Custom tools cost nothing to create and have no run limits, and any tool's output is capped at 1 MB. Retool's app building agent can also call tools on outside MCP servers connected as resources, while an app is built and when it runs, and people using the app can trigger only the MCP tools they have permission for; this is in public beta on cloud for all plans.

SourceRetool, docs.retool.com Agents Tools and changelog MCP in app builderread 2026-10-05

Omni is a conversational builder that sets up the work from a plain language request. It creates tables and fields, including field agents, builds interface pages and AI-generated interface elements, and creates or updates automations. A single agentic framework sits behind both Omni and Field Agents, and its agents reason, plan and orchestrate actions across the entire operation to accomplish complex tasks with minimal input. Underneath sits a dedicated automations engine as a core platform module.

Sourceairtable.com/platform/ai-agents, support.airtable.com using-omni-ai-in-airtable and the Airtable engineering blogread 2026-09-01

Agents choose which tools to call and in what order, invoke other agents as tools and talk to external agents over A2A, and agentic workflows combine Retool Workflows with agent steps. Each agent pairs a model and its tools with an orchestration layer that handles state, reasoning and planning.

It works in a ReAct loop, alternating between thinking and acting, with no fixed stopping point. A workflow calls an agent through an Invoke Agent block, and an agent used as a tool can run for up to eight hours. One business process can mix code that makes fixed decisions, model calls that make open ones, and people who weigh in.

SourceRetool, docs.retool.com Agents overview, Agentic workflows and Toolsread 2026-10-05

Field Agents can run automatically whenever data is added or updated, so they work continuously without anyone starting them. Automations provide event triggers, and work also starts from the web, desktop and mobile apps and through webhooks.

Sourceairtable.com/platform/ai-agentsread 2026-09-01

Agents start on their own when an email arrives at their unique agent address or when an outside agent calls them over the A2A trigger. People can also start them from the built in chat, from an Agent Chat component inside an app, or from an Invoke Agent block in a workflow. Chat threads can be shared with teammates.

SourceRetool, docs.retool.com Agents overview and Triggersread 2026-10-05

Agents are grounded in the customer's relational base and can read across connected data. Document analysis agents turn contracts, invoices and reports into structured fields, and web search agents pull real time external data into the same records.

Sourceairtable.com/platform/ai-agentsread 2026-09-01

Retool Vectors is a managed vector database for Retool AI. Retool extracts text from documents, splits it into chunks and computes embeddings with OpenAI's embeddings API, and it crawls URLs into new embeddings within published limits of three concurrent crawls, 25 URLs and a 15 second timeout. When an AI action references a Vector, the most relevant content goes into the model request. Agents have core tools that list Vectors, run a semantic search, fetch the chunks of a document or URL, and list the documents and URLs stored.

SourceRetool, docs.retool.com Retool Vectors quickstart and Agents Toolsread 2026-10-05

Agent outputs persist as structured fields in the relational base and become shared state that other agents read, and HyperDB scales to 100M records. There is no cross session agent memory or learned context layer separate from the data store.

Sourceairtable.com/platform/hyperdbread 2026-09-01

An agent keeps the context of its chat and its run as it calls tools, within one session. The orchestration layer manages memory and state during a run. Retool does not describe memory kept from one run to the next.

SourceRetool, docs.retool.com Agents overview and FAQread 2026-10-05

Admins enable AI per workspace, customers control what data agents can access, role based access control applies and all agent activity is auditable. There is no approval gate or review step before an individual agent action runs.

Sourceairtable.com/platform/ai-agentsread 2026-09-01

Agent tools run under Retool's role based access control and data level permissions, and anyone can watch an agent step by step as it works. Tools that need approval cannot be used over A2A, which does not support that waiting state. Retool lists consent before tool execution among its agent controls, and its launch post says approval of actions before they run "can be configured for every tool call."

In Retool's examples, an agent waits for a person's go ahead before it creates tickets or sends calendar invitations. The tool settings cover timeouts and typed parameters, and Retool does not show where the approval is switched on. Agents act on behalf of users, so they never use a tool or data source without permission.

SourceRetool, docs.retool.com Agent to agent and Agents overview, and retool.com blog on agent developmentread 2026-10-05

Airtable holds SOC 2 Type 2, with the report available on request, ISO/IEC 27001:2022 and ISO/IEC 27701:2019 with published certificates, TX-RAMP Level 2 with a certificate, and HIPAA. It also runs GDPR, UK GDPR and CCPA/CPRA programs, offers a signable DPA and keeps a subprocessor list.

Controls include SAML single sign on and extra admin features on Enterprise, two-factor authentication, customizable collaborator permissions, view access restricted by password or email domain, SCIM provisioning and enterprise key management that lets customers hold their own keys at rest.

Airtable also runs a HackerOne bug bounty program and publishes security and privacy whitepapers, and completed CAIQ, SIG Lite and HECVAT questionnaires are available under NDA.

Sourceairtable.com/company/trust-and-securityread 2026-09-01

Retool offers role based access control through administrator permissions, enforced two factor authentication and detailed admin audit logs on Business and Enterprise, and holds SOC 2 Type 2, with audit reports available through the Trust Center at trust.retool.com. Admins can sign out any user remotely, and every sign in is logged with its device type and IP address. Data is encrypted in transit and at rest, a vulnerability disclosure program takes reports from researchers, and outside security firms audit the platform regularly. The pricing page puts SSO and SCIM on Enterprise.

SourceRetool, docs.retool.com security practices and retool.com/pricingread 2026-10-05

Enterprise audit logs are available through the admin panel and through a dedicated Audit Logs API with published event types and several years of history. Revision history on each record gives a visual activity feed of every change, and that is where Field Agent output lands, so an agent's writes can be inspected at the row it wrote. Admin reports cover AI credit consumption, API and token owners and enterprise key management status, and all Field Agent activity is governed and auditable.

Sourceairtable.com developer audit log event types, airtable.com/company/trust-and-security and airtable.com/platform/ai-agentsread 2026-09-01

Each agent run shows token usage, estimated cost, total runtime and tool calls above a replay of the run with tool chips for each step, and any running agent can be watched step by step; admin audit logs sit on Business and Enterprise. Each run also lists its run ID, status, timestamp and duration.

Dashboards cover all agents or a single one, filter by time from the past 15 minutes to the past month, and stream events live. An agent graph shows in real time which resources an agent is touching. Thought logs show an agent's reasoning, with token, latency and cost figures for each run and in total.

SourceRetool, docs.retool.com Monitor agents and retool.com blog on agent developmentread 2026-10-05

Customers can have their data stored in the European Union, and Enterprise customers get enterprise key management, so data at rest is held under customer-owned keys. Models can also be served through Amazon Bedrock so providers never touch customer data. There is no self-hosted, on-premises, VPC or customer-cloud deployment and no FedRAMP authorization, so the platform is cloud only, with a residency option.

Sourceairtable.com/company/trust-and-security and airtable.com/company/data-residency-faqsread 2026-09-01

Retool runs as multitenant cloud on AWS or as 'Self-hosted services in your VPC/VPN', and Agents run on both cloud and self hosted instances, including email triggers through Mailgun on self hosted. Self hosting comes with the Enterprise plan. When self hosted, no Retool system stores customer data and no Retool staff can reach it, and only usage information and orchestration metadata go back to Retool. The cloud service is built to keep running through the loss of a server or a whole data center, and self hosted instances can switch A2A off with a feature flag.

SourceRetool, docs.retool.com security practices and Agent to agent, and retool.com/pricingread 2026-10-05

Customers can adopt packaged assets in three forms without building anything. AI Plays is a browsable library of premade AI workflows at airtable.com/ai-plays, each with its own page, including campaign concept ideation, brand and compliance review, localization of image and copy variants, and contract operations.

Four ready agent types are offered, document analysis, web search, image generation and custom. A catalog of pre-set field agents lets a user pick a suggested agent, which opens Omni with a request to build it, so the agent is adopted in one step and stays editable. An extensive template gallery and a public marketplace sit alongside.

Sourceairtable.com/ai-plays, airtable.com/platform/ai-agents and support.airtable.com using-airtable-ai-in-fieldsread 2026-09-01

Retool Agents 'offers several pre-defined templates' for specific use cases, but Retool names only one publicly, the Competition Scanner; the rest appear only inside the product. Competition Scanner comes loaded with mock competitor data from Retool Database and needs custom tools. The core tools are building blocks for agents, not agents in their own right.

SourceRetool, docs.retool.com Create from template and Agents Toolsread 2026-10-05

Enterprise customers decide which models are enabled and choose among OpenAI, Anthropic and Meta models. Anthropic Claude models are served through Amazon Bedrock so providers never handle customer data, and providers neither retain data nor train on it.

Sourceairtable.com/platform/ai-agentsread 2026-09-01

Builders 'choose your favorite LLM' and 'Choose from multiple AI models to balance performance and cost', on Retool managed keys for certain providers or their own key with a model provider. With Retool managed keys, the provider handles data under its agreement with Retool. With the customer's own key, the customer's agreement with that provider applies. Agent hours are priced by model, with more capable models costing more, and paid plans include up to 20 free agent hours a month.

SourceRetool, docs.retool.com Agents FAQ and retool.com/pricingread 2026-10-05

Developers get the Web API, Enterprise API, Audit Logs API and Shares API with a developer portal, personal access tokens, webhooks, scripting and custom extensions. An official MCP server and skills expose bases, records and workflows to external agents, and a Claude integration is promoted on its own landing page.

Sourceairtable.com/developers/web/api/introductionread 2026-09-01

Agents built in Retool are callable from outside over A2A, with an A2A trigger and an A2A endpoints reference, and the platform extends through custom tools, JavaScript, Python and SQL. Each agent has its own API key, sent in an X-Api-Key header and rotated as needed, and the endpoints accept HTTP with REST or JSON-RPC. Outside agents can send a message, poll for task updates, stream them over Server Sent Events or cancel a task. In the other direction, custom tools and resource queries can call outside agents over REST.

SourceRetool, docs.retool.com Agent to agent and A2A endpoints referenceread 2026-10-05

Omni produces a working configuration that the user reviews and then refines in conversation, and its limits are spelled out, including which interface layouts it cannot create. Automations keep run history for debugging a failed step. There is no agent evaluation harness, scored or repeatable test run, regression suite or output scoring.

Sourcesupport.airtable.com using-omni-ai-in-airtable and airtable.com/platform/automationsread 2026-09-01

Evals test agents systematically: builders create datasets and test cases, run evals and compare runs side by side, and 'fine-tune with built-in evals'. Each test case holds an input and the expected output. Tool choice cases check that the agent picks the right tool with the right parameters, and final answer cases grade the reply. Programmatic reviewers check exact matches, valid JSON or XML, schema matches, contained strings or Levenshtein distance, and LLM judges score tone, factuality or closed questions, each on a scale from 0 to 1. Two runs can be compared at a time.

SourceRetool, docs.retool.com Agents Evalsread 2026-10-05

Web search agents bring real time web data into records, but browser control, page navigation and computer use are not described anywhere in the product.

Sourceairtable.com/platform/ai-agentsread 2026-09-01

Agents act through core tools, custom tools, queries, workflows, other agents and MCP servers; the web search core tool is a search, and Retool offers no browser, desktop or computer control. A second web tool reads a page's content, and the code tool runs JavaScript or Python, but none of the tools clicks through a site or works a screen.

SourceRetool, docs.retool.com Agents Toolsread 2026-10-05