Why Marshal
Production agents need more than a model call. They need shared primitives: prompts and skills that stay versioned, memory that survives sessions, tools that stay vaulted, and gates that match your org’s bar. Marshal is that harness—one toolkit across frameworks and channels.
The operating reality
Teams stitch gateways, prompt stores, skill folders, memory services, and approval flows separately. Each agent becomes a one-off. Marshal consolidates the agent primitives under one org: compose once, run from SDK, CLI, console, or MCP, and promote with the same evals and guardrails.
What Marshal is for
| Primitive | Job |
|---|---|
| AI Gateway | Route LLM traffic with auth, cache, rate limits, and failover |
| Prompts & skills | Versioned instructions and procedures without redeploying every agent |
| Memory & context | Durable facts and live estate context shared by agents and operators |
| Tools & MCP | Connectors that publish governed tools to frameworks and IDEs |
| Evals & guardrails | Quality gates and spend / policy caps in the execution path |
| HITL | Explicit human approval where judgment is required |
| SDK & CLI | Same APIs for Python, TypeScript, and automation |
Design principles
- Harness engineering — Build from primitives, not ad-hoc scripts. See Harness Engineering.
- Any framework, same toolkit — Adapters for major agent runtimes; your orchestrator stays yours.
- Operate without redeploy theater — Update prompts, skills, and memory; agents pick up the next turn.
- Govern in path — Gateway limits, guardrails, and HITL sit beside the work—not in a slide deck after the fact.
- Pair with the IDE — Relay covers coding-agent actions; Marshal covers production agent operations. Full Platform covers both.
Who adopts Marshal
- Platform and AI engineering teams standardizing how agents are built and run
- Orgs that want one MCP and gateway surface for connectors and models
- Teams promoting prompts and skills with evals before traffic shifts
- Companies that already use Relay and want the same accountability for backend agents
What you get next
- Create an org
eis_*key — Tokens and API keys - Follow Getting started — connectors, prompts, skills, gateway
- Run a live sample — Live SDK examples
Product hub: exemplar.dev/marshal · Pricing: exemplar.dev/pricing .
Last updated on