Open Source · Pre-GA · Apache 2.0
Build, host, and
orchestrate tool-using
AI agents.
Agentrail gives you a composable runtime, a hosted server layer, multi-agent orchestration, and a clear extension model without locking you into any platform.
import { defineAgent } from "@agentrail/core"; import { defineProfile } from "@agentrail/app"; export const researchAgent = defineAgent({ id: "researcher", model: { provider: "anthropic", modelId: "claude-sonnet-4-5", apiKey: process.env.ANTHROPIC_API_KEY, }, system: "You are a research assistant.", }); // stream a response const result = await researchAgent.stream( "What are the latest advances in agentic AI?", );
Why Agentrail
The right level of abstraction.
Most teams land between two extremes: scripts that break at scale and platforms that create lock-in. Agentrail is designed to sit in between.
Too little structure
Ad hoc scripts
- No session management
- Tool contracts are implicit
- Orchestration is bespoke every time
- Testing is painful
- Hard to extend without rewriting
Just right
Agentrail
- Composable runtime with clear contracts
- Host layer for chat and stream lifecycles
- First-class multi-agent orchestration
- Plugin and context provider extension model
- No platform lock-in, fully self-hosted
Too much lock-in
Hosted platforms
- Opaque internals
- Metered billing surprises
- Limited extensibility
- Vendor controls your data
- GUI-first, not code-first
What's included
Everything you need to ship agents.
A focused set of packages. Each has a clear responsibility and can be used independently.
@agentrail/core
Agent Runtime
Define agents with typed tool contracts. The LLM loop, tool dispatch, provider abstractions, prompt SDK, and usage tracking sit in one stable core.
@agentrail/capabilities
Capabilities
Sandboxed filesystem and browser tools, knowledge-base retrieval, skills registry, and multi-agent orchestration. They can be mounted with defineProfile({ capabilities: [...] }).
@agentrail/app
Host Layer
createAgentApp mounts chat and stream endpoints in one call. defineProfile wires model, prompt, and capabilities. Session management, plugins, slash commands, and typed config loading are included.
@agentrail/capabilities · orchestration
Multi-Agent Orchestration
Spawn sub-agents, send typed work, wait on conditions, and recover from failures. State can persist across restarts through JSONL-backed mailboxes.
@agentrail/app · SessionManager
Session Management
Filesystem-backed session storage with append-only JSONL history, automatic context compaction, and conversation branching.
@agentrail/capabilities · sandbox
Sandboxed Execution
Docker-based isolated execution for browser automation, shell commands, and file I/O. Execution is contained per session.
Architecture
Layered by design.
Start with the recommended SDK path, then drop to lower-level primitives only when tighter control is needed.