Tray.ai

Agent Gateway exposes Tray workflows as composite tools and individual operations from 700+ connectors as tools over MCP, with tools authenticating by shared service credentials or the end user's own, an access allowlist, OAuth2 or API token for the calling client, RBAC and rate limits; agents built in Merlin Agent Builder take actions by running Tray workflows as tools.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-gateway/overviewread 2026-09-22

Tray is a workflow platform with an agent layer on top: workflows sequence steps with conditional logic, branching and looping, retries and timeouts, and an agent project combines a scope, data sources, workflows as tools and a model, so autonomous agent steps run alongside deterministic workflow nodes. Composite tools package multi-step business logic as a single callable tool.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-builder/overviewread 2026-09-22

Adding a data source in Merlin Agent Builder connects a system (Google Drive, Gmail, Confluence, SharePoint or a custom connector), syncs and processes the documents and stores the content in a vector table that supports semantic search, which the agent queries through a Knowledge Search tool included by default. New customer knowledge joins a maintained index without retraining, one that persists, scales past the context window and stays queryable.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-builder/data-sourcesread 2026-09-22

A person sits in the path at both layers. In workflows, an approval step can be inserted anywhere and routed to Slack, email or a custom form, and the run pauses and resumes automatically once the decision is made; agents built in Merlin Agent Builder act by running those workflows as tools.

For agents directly, Tray states that guardrails pause an agent when confidence is low or when an action falls outside its defined scope, and surface the decision to a person before it continues, and the Merlin Agent Builder page lists approval workflows for sensitive actions. Around that sit the constraints: an Agent Gateway access allowlist, RBAC, rate limits, agent scope, log masking and an execution kill switch. So a person can approve before an action commits, inside the channels a team already uses.

Sourcetray.ai/glossary/human-in-the-loopread 2026-09-25

SOC 1 and SOC 2 Type 2, HIPAA, GDPR and CCPA are stated across the platform, and Tray says every agent action, MCP tool call and workflow run is traceable to a user. The access surface is documented in the enterprise docs: organizations and workspaces, user roles, two-factor settings, an access allowlist and RBAC on Agent Gateway, log masking to hide a step's inputs and outputs, and administration of which Merlin AI features are enabled.

Sourcetray.ai/platformread 2026-09-22

Every workflow run produces debug logs showing each step's input and output, agents have their own logs and debugging view, Agent Gateway has an observability and monitoring page for tool execution, and Tray streams workflow logs to an external endpoint for the customer's own tooling, with log masking, disabling of log storage and anomaly detection as controls. The platform page states every agent action, MCP tool call and workflow run is traceable to a user.

Sourcetray.ai/documentation/platform/enterprise-core/logs-debugging/log-streamingread 2026-09-22

Tray states that Merlin Agent Builder includes short and long term memory, so agents keep track of session history and refer back to a user's prior conversations automatically, with a sliding context window keeping each exchange focused. Tray frames the problem it solves as an agent forgetting a user's earlier issue every time that user returns, which places the memory across conversations rather than inside one. Knowledge synced into the agent's vector table is a separate retrieval structure over documents.

Sourcetray.ai/company/news/merlin-agent-builder-user-adoption-releaseread 2026-09-25

A Tray instance is hosted in a data region the customer picks, US, EU or APAC, so business data does not leave the region it originated in, with region-specific API base URLs and multi-region setups for multinational or embedded use. For systems behind the firewall, Tray documents an on-premise agent and AWS PrivateLink connectivity.

Sourcetray.ai/documentation/platform/enterprise-core/organisation-management/regional-hostingread 2026-09-22

Buyers start from preconfigured agent accelerators, each with a predefined scope for its use case: a Knowledge agent, an ITSM agent and a Support ticket agent, alongside prebuilt tools and workflow templates that can be customized.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-builder/configurationread 2026-09-22

Work reaches Tray without a person asking: workflows start from connector triggers and webhook subscriptions, scheduled polling for new data and an alerting trigger, and agents are reachable through interaction channels in Slack and Microsoft Teams as well as the API.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-builder/interaction-channelsread 2026-09-22

An agent's model is a configuration choice: Tray ships a native default model and documents bringing your own by connecting API keys from third-party providers such as OpenAI or AWS Bedrock, choosing the provider and the specific model version in the agent dashboard.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-builder/configurationread 2026-09-22

Tray publishes OpenAPI specifications for the Tray Platform and Tray Embedded APIs, a Connector Development Kit for building and testing custom connectors, developer tools including a connector tester and operations explorer, embedded auth dialogs for the customer's own end users, and Tray Headless MCP for connecting Cursor, Windsurf, Codex or Claude Code to the platform.

Sourcetray.ai/documentation/platform/tray-headless/tray-headless-mcpread 2026-09-22

An AI evaluation framework sits in Tray's documentation library: a template that uses an LLM as a judge to review each input prompt and agent response against set criteria, covering response quality, conversation flow, edge cases and behavioral consistency. The judge returns structured feedback and quantitative metrics tracked over time, so a team can compare versions of an agent and see whether a change improved it. That sits beside the built in test chat, which shows each reply with the tools considered, whether the knowledge base was searched, processing time and token usage.

Sourcetray.ai/documentation/library/template/6480885f-d884-469c-8c80-77a2f36a4b61-get-started-with-ai-evaluationsread 2026-09-25

Tray acts through connectors, workflows and MCP tools that call APIs; no browser, desktop or remote computer session that Tray drives for an agent is documented. No headless fetch or third party browsing engine of Tray's own is documented either.

Sourcetray.ai/documentation/platform/artificial-intelligence/agent-gateway/overviewread 2026-09-22