Technical Reference · Concepts
Concepts
A map of the core concepts behind the platform, from the sources tools come from through the servers that serve them and the controls that govern them.
These pages describe the core concepts of the platform and how they fit together. Sources are deployed to generate tool definitions, selected tool definitions are served to agents as MCP servers, and a layer of identity, observability, and policy controls governs what those agents can do.
Concepts with a dedicated page under this section are listed first. The rest link to the page that covers them in depth.
Building tools
Section titled “Building tools”The inputs that describe available functionality: OpenAPI documents, Functions, and external MCP servers. Uploading a source is the starting point for creating tools.
Immutable snapshots of a project’s sources and the tools generated from them. Every source change produces a new deployment, and the most recent successful deployment serves the project’s tools.
The output of processing a source: one definition per API operation, function tool, or external MCP tool, containing the metadata an LLM needs and the configuration the platform uses to execute the tool.
Overrides for a tool’s name, description, tags, annotations, and confirmation prompts, a prompt-engineering layer that improves how LLMs select and use tools without changing the underlying source.
Serving tools
Section titled “Serving tools”How tools reach agents. Each server exposes a selected set of tool definitions over streamable HTTP, with its own authentication, visibility, and team access.
An MCP server a third party already runs, registered by URL or from the catalog. Nothing is copied or re-hosted. The platform proxies whole sessions to the upstream and adds authentication, per-tool access control, metering, and logging in front.
Named sets of variables that hold the secrets and configuration sources need to reach upstream systems. Binding an environment to a server keeps credentials centralized instead of scattered across individual client configurations.
Distributable bundles that package MCP servers and skills together, assign them to roles, and publish to the Claude Code, Cursor, and Codex marketplaces through GitHub.
Identity and access
Section titled “Identity and access”Who is calling a server. The platform acts as the OAuth authorization server for a server’s endpoint, so every tool call carries both a subject, the person behind the agent, and a client ID, the agent itself.
- What an MCP client experiences
- Dynamic client registration
- Session duration and token lifetime
- Revoking access
Observing and securing
Section titled “Observing and securing”Captured agent conversations, each holding a full transcript of messages and tool calls with cost and token attribution. Sessions are the record investigations start from, and risk findings attach to specific messages within them.
MCP servers discovered in agent traffic that the platform does not manage, meaning servers members connected on their own. Every routed tool call records which server it resolved to, and calls resolving anywhere other than a platform-hosted host are flagged.
The rules that scan agent sessions for secrets, sensitive data, and prompt-defined risks. A policy binds detection rules, the scope of content they examine, the action taken on a match, and the audience it applies to.