Coresource Docs

Build and run complete AI agents. You describe the instructions, model, capabilities, and outputs. The platform runs durable work, stores files and artifacts, and exposes status, results, and events.

Get started#

QuickstartInstall the SDK, set your API key, and run your first two agents in about five minutes.AuthenticationAPI keys, environment variables, supplying your own fetch, and pointing the SDK at another deployment.Core conceptsAgents, runs, events, artifacts, and what durability actually guarantees.Agent typesOne shot, Single Agent, and Saga: what each is for and how to choose.

Browse by topic#

Running agentsStreaming, cancellation, files, artifacts, human input, and the error taxonomy.ToolsHosted tools, the runtime catalog, and writing handlers that run in your own process.SkillsPackaged procedures an agent can discover and invoke, published or local.MCPRemote tool servers, hosted connections, OAuth apps, and credential management.OrchestrationSaved agents, and composing them into sequential and parallel workflows.HTTP APIThe hosted agent resources for any language, including the OpenAI Responses-shaped endpoint.

Choosing an agent type#

The three types expose different execution capabilities. Choose the smallest type that supports the work.

Type	Reach for it when	What it cannot do
`one-shot`	The model needs to answer once. Summarising, extracting, rewriting.	No tools, no files, no iteration.
`agent`	The work needs tools, files, skills, or MCP, and finishes in a bounded number of turns.	No long-horizon planning; it iterates rather than plans.
`saga`	The work benefits from an explicit plan and independently executable steps.	It is not intended for a single model response.

A One shot agent returns one model response. A Single Agent runs an iterative tool-use loop until it finishes or reaches maxTurns. A Saga creates a plan and executes its steps durably, with independent steps eligible to run concurrently.

What the platform manages#

You supply the agent definition and the prompt. Coresource supplies:

Execution. Durable agent and saga runs continue server-side if your process exits or its network connection drops. The id returned by createAgent().start() is also a durable run id that another process can retrieve through the HTTP API. A run that needs a custom tool or local MCP server still needs that SDK process whenever it calls the local capability.
Credentials. Model-provider credentials and credentials for hosted MCP connections are managed by the platform. Credentials used by custom tools and local MCP clients remain in your process.
Storage. Input files and output artifacts are stored and served for you. Storage locations and runtime filesystem roots are not part of the public API.
Streaming. Durable runs expose live events and retained event history. One-shot responses return synchronously and do not expose a run stream.

Your process supplies one thing the platform cannot: the handlers for your own tools. When you pass a function as a tool, the model's call is routed back to your machine, your code runs there, and the result returns to the run. That is what makes it possible to give an agent access to your database, your internal API, or your browser automation without handing over any credentials.