A standalone runtime for portable agents

Define an agent once, then run it locally, inside your application, or on remote compute. Clear Ideas Agent Runtime is distributed as npm packages and does not require a hosted Clear Ideas service.

npm install @clearideas/agent-runtime

The package will become installable when the prepared npm release is made public; using the published runtime will not require a source build.

Quick startRun your first agent with one model API key.Build agentsLearn prompt chains, routing, loops, tools, and approvals.Embed Agent RuntimeCompose the TypeScript API inside an application.Remote executionRun agents through HTTPS workers or remote compute adapters.

How it fits together

Agent manifest
Reusable definition

→

Agent run manifest
Reference and inputs

→

Agent Runtime
State, models, tools, and execution

→

Adapters
Storage and compute

The agent manifest defines the agent. An agent run manifest references that definition and supplies invocation values. Host configuration selects model providers, credentials, and connections. Adapters connect persistence, compute, sandbox providers, and telemetry.

Authorization policy, sandbox contracts, durable run state, and execution engines are native parts of Agent Runtime. Agent manifests may narrow host-authorized connections and tools, but cannot grant themselves credentials or broader access. Docker and Modal sandbox providers ship as packages in the same project.

Choose a path

Goal	Start here
Try Agent Runtime locally	Quick start
Learn the agent patterns	Build agents
Use local or hosted models	Models and providers
Add MCP or application tools	Connections and tools
Add Agent Runtime to a service	Embed Agent Runtime
Choose packaged integrations	Adapter catalog
Run on ad hoc compute	Remote execution
Prepare a hosted deployment	Production guide

Execution model

• Agent manifests are YAML or JSON documents containing serializable values.
• Model, persistence, compute, sandbox, and telemetry integrations use adapters.
• Checkpoints record committed state and nested progress.
• Agent runs are sequential by default and can fan out dependency-safe prompt steps.
• Stateful steps and tool calls remain sequential.
• Runtime policy and host configuration authorize credentials, connections, tools, network access, and compute.