Twin

More than 5,000 apps connect out of the box through native integrations, agents can connect to any API, and where no connector exists Twin writes the integration itself. Agents read and write in tools such as Gmail, Google Sheets, CRMs and Slack.

Sourcetwin.soread 2026-09-27

Twin turns a goal into an agent that plans and executes the steps. The workspace Orchestrator creates, modifies and coordinates several agents that share context, proposing a plan the user reviews before anything runs.

Sourcedocs.twin.so/orchestratorread 2026-09-27

Agents use web search, deep research and the tools they connect to per run, and workspace agents share context. There is no maintained index over the customer's own documents.

Sourcedocs.twin.so/workspacesread 2026-09-27

Approvals are a named agent feature beside schedules and webhooks (an agent drafts a review reply and posts it after the user's OK). Browser automation opens a permission dialog showing the plan for the user to approve. The Orchestrator's plan is reviewed before agents are built.

Sourcedocs.twin.so/web-agentread 2026-09-27

Twin says it is audited and certified without naming a framework, and it links a trust center on app.bastion.tech, but it publishes no named attestation and has no security page. Access control covers only workspaces, API keys and per member credit caps, with no roles or SSO named.

Sourcetwin.so/pricingread 2026-09-27

Each run keeps its steps for review (which pages the agent visited and what it captured). The REST API exposes run events, webhooks fire on run.completed and run.failed, and usage tracking shows credit consumption per member and agent.

Sourcedocs.twin.so/rest-apiread 2026-09-27

Run Mode uses the agent's memory from previous builds, and agents in one workspace share context. Twin does not give that memory a scope, a lifetime or a way to review it. Login sessions carried between runs are browser state, not agent memory.

Sourcedocs.twin.so/build-vs-runread 2026-09-27

Twin runs agents only in its own cloud, on isolated cloud computers per agent, with nothing to deploy. There is no region choice, residency option or self hosting.

Sourcedocs.twin.so/guidelinesread 2026-09-27

The Employee Library (beta) holds more than 200 ready-made agents shared by the community that a customer can hire as is, such as B2B Prospect Hunter, Niche Content Composer and SaaS Founder Signal Scout, each doing its own job. Agents can be shared publicly or sold through a store.

Sourcetwin.so/agentsread 2026-09-27

Event based triggers start agents when something happens in a connected app (a new message or record), time based triggers run them on a schedule, and the REST API starts them from the customer's backend.

Sourcedocs.twin.so/triggersread 2026-09-27

The user picks high or low reasoning before a build, and runs use a lighter model. No model provider is named, and there is no provider choice or bring your own model route, so the choice is a cost tier.

Sourcedocs.twin.so/build-vs-runread 2026-09-27

A REST API at build.twin.so with an OpenAPI 3.1 schema and x-api-key authentication creates agents, triggers runs, inspects events and manages API keys.

Sourcedocs.twin.so/rest-apiread 2026-09-27

To confirm an agent worked, the user reviews its run steps and outputs, and agents are adjusted through chat. That is debugging after a run, and there is no evaluation harness, test cases or release gate.

Sourcedocs.twin.so/troubleshootingread 2026-09-27

The Web Agent controls a browser on an isolated cloud computer dedicated to the agent, handling logins, multi step flows and dynamic pages when a service has no API or the API lacks a scope. Twin calls it the costliest and least reliable mode and recommends APIs first.

Sourcedocs.twin.so/web-agentread 2026-09-27