SuperAnnotate

Agents connected in Agent Hub can carry up to 20 function tools and, on OpenAI, Anthropic and Gemini models, a web search tool, and a function call comes back in the tool_calls field of the agent's JSON response for a following pipeline step to act on. Inside the platform, the MCP server lets pipeline agents read projects and items and write annotations and statuses. Outside systems are reached through Orchestrate's webhook and email actions and custom Python actions that call any API with stored secrets, while storage integrations hold read only access to the customer's buckets.

Sourcedoc.superannotate.com/docs/agent-hub-overviewread 2026-11-09

Orchestrate pipelines chain events, actions and agent nodes on a canvas, and each node's output becomes the context payload of the next, which an agent prompt can address field by field with JSONPath. Nodes carry success and failure branches, custom Python actions run with their own libraries and secrets, and a router can approve confident results automatically while sending the rest to expert review. A rerun replays an action with its original payload and runs every node after it, and whole pipelines export as templates for other teams to import.

Sourcedoc.superannotate.com/docs/about-orchestrateread 2026-11-09

Agent Hub connections hold instructions, model settings, tools and an output format, with no knowledge base, document index or vector store for an agent to retrieve from. Context reaches an agent as the payload of the event that started the run, values pulled into the prompt with JSONPath, files attached to a request, and item annotations read through the MCP server at run time. Embedding generation and natural language search were removed from Vector and Text projects, and retrieval augmented generation appears on the platform as a system to evaluate rather than a store agents query.

Sourcedoc.superannotate.com/docs/agent-hub-overviewread 2026-11-09

Human review sits inside the same pipelines agents run in. A model can prelabel or score each item, low confidence results go to an expert review checkpoint that approves them into the dataset or sends them back, and items an LLM judge scores below a threshold go to two evaluators with disagreements escalated. Custom roles carry approve and disapprove rights and item statuses gate what moves on. Administrators can deselect MCP tools an agent should not use, though MCP tool calls run without approval by default.

Sourcesuperannotate.com/agent-hubread 2026-11-09

Organization owners set up SAML 2.0 or OIDC single sign on with Azure AD, Okta, Ping Identity and other providers and can enforce it for every member, on the Pro and Enterprise plans. Team Owner, Team Admin and Contributor roles, project roles, custom roles and group data scopes decide who can connect models, reach Orchestrate, manage API keys or download data. The trust center lists SOC 2 Type II, available on request, and an ISO/IEC 27001:2022 certificate open to download, alongside annual penetration testing and encryption in transit and at rest.

Sourcedoc.superannotate.com/docs/org-settingsread 2026-11-09

Monitoring lists every action and agent run across a team's pipelines with its status, triggering event, project and creation time, and charts them by count or compute time over windows of up to six months. Opening an agent run shows its parameters and prompt, the chain of triggers that led to it, its input, its output and a trace of each MCP tool call with request and response payloads, kept for 14 days. Telemetry adds logs of every annotation status change, working time and assignment history, which reporting APIs export as CSV.

Sourcedoc.superannotate.com/docs/pipelines-monitoringread 2026-11-09

Agent requests go straight to the model provider's endpoint and come back as received. A pipeline passes one node's output to the next within a single run, and a rerun replays an action with the payload it first received, so nothing an agent concludes carries into later runs. Items, annotations and statuses persist because they are the project's data, which an agent reads or changes through the MCP server like any other record. Agent Hub connections keep configuration such as instructions and model parameters, with no memory store scoped to a user, team or workflow.

Sourcedoc.superannotate.com/docs/run-agents-functionread 2026-11-09

Data can sit in SuperAnnotate's US cloud, its EU cloud or an on premises installation in the customer's own environment, and items can stay in the customer's AWS, Google Cloud or Azure storage, where the platform holds read only access and only displays them. Agent Hub can call custom model endpoints on the customer's own infrastructure. The privacy policy places the hosted service on servers in the United States, and AWS in the USA is the only cloud host on the subprocessor list.

Sourcesuperannotate.com/solutions-agentsread 2026-11-09

New projects can start from a library of form templates such as chat rating, side by side model comparison and image generation feedback, each previewed with its own README before import, and an assistant in the editor drafts an interface from a description. No template lays out an agent run step by step, and Agent Hub ships no prebuilt agent or packaged judge, so teams assemble agent review forms and judges themselves. Pipeline templates are JSON exports of a team's own pipelines, and the benchmark datasets are data products rather than agents.

Sourcedoc.superannotate.com/docs/mm-templates-use-casesread 2026-11-09

Pipelines start on their own from platform events: a project, folder, item or export being created, updated or deleted, or a contributor joining or leaving a project. An incoming webhook lets any outside system start a run, and a periodic event fires hourly, daily, weekly or on a custom interval in a chosen time zone between set start and end dates. Custom events let a team run a pipeline by hand, and an agent node receives the whole payload of the step before it appended to its prompt, so the triggering event is its starting context.

Sourcedoc.superannotate.com/docs/orchestrate-eventsread 2026-11-09

Each Agent Hub connection names its own provider and model, chosen by the team from OpenAI, Anthropic, Gemini, Fireworks AI, Vertex AI, Databricks and AWS Bedrock, or a custom endpoint for an internal or on premises model, and runs on credentials the team stores. Teams set reasoning effort, verbosity, temperature, token limits, timeouts and retries per connection and override them per pipeline node, with up to 200 connections a team, so judges and labeling agents can run on different models at once.

Sourcedoc.superannotate.com/docs/agent-credentialsread 2026-11-09

The Python SDK, installed from PyPI, works with the platform's own projects, folders, items, annotations and subsets, authenticated with team, personal or organization API keys that can be generated, rotated and revoked, and reporting APIs export activity data. A hosted MCP server takes a bearer team or personal key and exposes ten tools that list and read projects, folders, items and annotations and set item annotations and statuses, usable from pipeline agent nodes or an outside MCP client.

Sourcedoc.superannotate.com/docs/superannotate-mcp-serverread 2026-11-09

Agent runs import with their tool actions, decisions and responses into a review interface the team shapes to its own agent, laying out each step from input and reasoning to tool call and response. Reviewers score a run on dimensions such as tool use, reasoning and task fit, correct a wrong tool argument in place and turn reviewed runs into datasets, while LLM judges can score first and pass weak results to experts. Analytics track agreement between each evaluator and the judge, and an audit trail records every score change and approval.

Sourcesuperannotate.com/solutions-agentsread 2026-11-09

Agents on the platform work through model calls, function tools, web search and the MCP server, and no agent drives a browser, a desktop or another application's screen. RL environment packs for browser and computer tasks are built with SuperAnnotate's research team and expert services rather than run from the platform, and the computer use trajectories it sells are recorded datasets for training and evaluating other agents. The multimodal editor runs in a browser, operated by the people doing the review.

Sourcedoc.superannotate.com/docs/agent-hub-overviewread 2026-11-09