TinyCommand

Any action from the 458 connected apps can become an agent tool: a builder wraps the action on the canvas, saves it as a tool and adds it to the agent, which fills the inputs at run time, so an agent can send a Gmail, post to Slack, update HubSpot or add a row.

Custom tools also wrap any REST API with headers, input parameters and output mapping, or a published workflow with a manual trigger, and agents read and write the workspace's own TinyTables, look up people and companies, search the web and read documents. App connections are authorized once per workspace and shared across members' workflows and agents.

Sourcetinycommand.com/docs/agents/toolsread 2026-10-18

Given a goal, an agent makes a plan, picks a tool, reads what comes back and decides the next move, repeating until the job is done, and it can call prebuilt sub agents for research, writing, search and enrichment.

The Agent node runs a full agent as one step on the workflow canvas, next to more than 60 step types: if and else branches, for each, loop until, repeat, jump to, match pattern, delays, merges, aggregators and formulas. Nodes can continue on error, and loops with delays build retries and exponential backoff. A plain sentence lays out a starter workflow that opens on the canvas for editing before it is published.

Sourcetinycommand.com/docs/workflows/actions/agent-noderead 2026-10-18

Each agent has a knowledge base of uploaded PDF, Word, plain text and CSV files up to 10 MB each, which are chunked and embedded for search, and before answering the agent retrieves the most relevant passages across all files and grounds its reply in them, naming the source file.

Built in research tools add grounded web answers with sources through TinySearch, and workflow steps crawl a site, map its URLs, scrape a page or extract structured data from it. Vision steps read uploaded documents and images, so an application form's attachments can be parsed mid flow. File processing status shows when each upload becomes searchable.

Sourcetinycommand.com/docs/agents/knowledge-baseread 2026-10-18

A human in the loop node pauses a workflow and sends a branded review page with instructions, the data to review, optional file attachments and at least two buttons, using templates for approval, data review or manual input, and the person reviewing can edit the content when allowed.

Each request carries a timeout of 24 to 72 hours by default and a fallback that picks a button automatically. Agents can stop for sign off before acting through a Human Approval tool, and handoff rules pass a conversation to a person on an explicit request, low confidence, a sensitive topic, repeated answers, a sentiment drop or too many turns, with the full transcript.

Sourcetinycommand.com/docs/workflows/logic/hitlread 2026-10-18

Workspaces use Owner, Admin, Editor and Viewer roles, folder and file permissions on paid plans and two factor sign in, and each workspace is fully isolated. API keys are scoped to the whole workspace or to tables, workflows, forms or agents, set read only or read and write, and expire after 30 days, 90 days, a year or never.

Webhooks get unguessable URLs, IP filtering and HMAC signatures, credentials are encrypted apart from application data, third party penetration tests run quarterly and a DPA is offered. There is no single sign on, and SOC 2 and ISO 27001 certification are still in progress.

Sourcetinycommand.com/securityread 2026-10-18

Execution history lists every workflow run with status, trigger time, duration and credits used, and opening a run shows the node timeline, the exact input and output of each node and detailed errors for any that failed, with export for audit and retention of 15 days on Free and Basic and 30 on Professional and Agency.

An agent's Activity tab shows each conversation with the user's messages, the agent's replies, every tool and sub agent call with its input and output, and the latency of each response. An audit log available to owners and admins records logins, failed logins, two factor changes, API keys and changes to workflows, forms and connections, with user and IP address.

Sourcetinycommand.com/docs/workflows/running/execution-historyread 2026-10-18

Memory is a setting on the agent that keeps what came before across a chat and across runs, and inside a workflow an agent can keep one conversation going across several steps of the same run. Agents also read and update TinyTables, so customer records persist as ordinary data that later runs can look up. TinyCommand gives no retention period for agent memory and no screen where a person views, edits or clears what an agent remembers, so in practice control over long lived context runs through the agent's instructions and the tables it writes.

Sourcetinycommand.com/tinyagentsread 2026-10-18

The service runs on AWS with databases in two regions, the United States and Europe, and customers with European residency requirements can have their data hosted in the European region to support GDPR and UK GDPR, with the region assigned by TinyCommand to fit those requirements.

Vision and agent features run on models TinyCommand hosts itself, so data those features process is not sent to third party AI providers unless a customer connects its own provider key. There is no self hosted or private cloud edition, backups are encrypted daily, and data stays exportable for 30 days after an account closes.

Sourcetinycommand.com/securityread 2026-10-18

Six prebuilt agents are ready to run on their own or as sub agents: TinyScout researches a company or prospect and returns a summary, intent score and fit read, TinyComposer writes personalized messages and emails from a recipient and a goal, TinySearch answers from the live web with sources, and Company Enricher, Person Enricher and Email Finder fill out records from a name, domain or email. Recipes supply prebuilt systems, some installable in place, alongside workflow templates, form templates and quick start trigger templates, and the review node ships approval, data review and manual input templates.

Sourcetinycommand.com/docs/agents/capabilitiesread 2026-10-18

Workflows start from seven trigger types: an app event such as a Slack message, Stripe payment or GitHub pull request, a TinyForms submission, a row created, updated or deleted in a TinyTable, a date column reaching a threshold, an incoming webhook, a cron schedule and a manual run.

Agents go live at a public URL and embed on any site as inline, full page, popup, corner button, floating or sidebar chat, with optional voice input and file attachments, and the same agent runs when a workflow reaches its Agent node. An API key scoped to agents lets outside applications hold agent conversations, and workflows can be triggered over the REST API.

Sourcetinycommand.com/docs/workflows/triggersread 2026-10-18

Each agent picks its provider, model, temperature and maximum response length on its Brain tab, choosing among OpenAI, Anthropic, Google Gemini and Mistral models with the customer's own key, or TinyGPT, the built in default that needs no key and draws on credits.

Guidance on the tab matches model to job, from a small fast model for a general support bot to a larger one for multistep reasoning or long document questions. The choice is made once for each agent, so different agents in one workspace can run on different makers' models, and enterprises can point an agent at their own Azure OpenAI deployment.

Sourcetinycommand.com/docs/agents/brainread 2026-10-18

A REST API at api.tinycommand.com lists tables and reads, creates, updates and deletes rows, triggers workflows, reads form submissions and holds agent conversations, authenticated with bearer keys scoped by product and permission and set to expire. Incoming webhooks start workflows, HTTP request steps and custom API tools reach any outside endpoint, and published workflows become callable tools for agents. An MCP server with 14 tools for building systems, operating tables and running agents is documented, but it sits behind a waitlist while authentication and rollout finish.

Sourcetinycommand.com/docs/account/api-keysread 2026-10-18

Agents are tested in live chat before going live against a set method: happy path questions, edge cases such as ambiguous or off topic messages, safety prompts that try to override instructions, and each tool fired with the right parameters.

Every agent deployment creates a version that can be compared with any other across instructions, knowledge, tools, identity and sub agents and restored as a new version, and workflows keep the same history with a canvas view of each past version. Workflow runs can be tried on a sample and read step by step before publishing. There is no automated suite of scored test cases or regression runs after a change.

Sourcetinycommand.com/docs/agents/testingread 2026-10-18

Agents come with built in tools that can be switched on one by one, among them web browsing, code execution and file processing, and workflow steps crawl a site, map its URLs, scrape a page or extract structured data from it, with Web Scrape also offered as a building block for any build. Those tools read and extract web content during a run. No agent clicks through a site, fills in a form, signs in to a portal or controls a desktop, and there is no recording of on screen actions, so a lender that needs data entered into a broker portal still does that step by hand or through an API.

Sourcetinycommand.com/docs/agents/capabilitiesread 2026-10-18