- CollectionAgent Runtime & Infrastructure
- Classificationagent-native
- StatusOpen source · early-stage · Jan 2026 launch
- MCPPlanned
- Agent SkillUnavailable
- Interfaceswebhook · SDK · http
What it does
Polos is an opinionated runtime for AI agents that bundles four concerns into one stack:
- Sandboxed execution — agent code runs in a Docker or E2B container with a clean, throw-away filesystem and a built-in tool kit (exec, files, search).
- Durable workflow — each step is checkpointed; on failure, the runtime retries with prompt cache hits to avoid re-paying LLM costs.
- Multiple triggers — HTTP, webhook, cron, and event triggers are first-class; agents are deployed once and react to many sources.
- HITL via Slack / UI — pause-and-approve gates without ad-hoc plumbing.
The pitch is that it sits in the gap between code-execution sandboxes (E2B) and durable workflow engines (Trigger.dev, Inngest): you get both, agent-shaped, in one runtime — at small-team operational cost.
Agent-facing interfaces
- Catalog protocols: webhook, SDK, http
- MCP: ⚠️ Not yet published as a first-party MCP server. Polos consumes MCP tools (it can call any MCP server from inside the sandbox); a server-side MCP surface is on the roadmap.
- Agent Skill: ⚠️ Not yet published.
- Onboarding type:
sdk
How agents onboard
# Install:
pip install polos
# or
npm install polos
# Run an agent inside the runtime:
import polos
@polos.agent(triggers=["http", "cron", "webhook"])
def my_agent(ctx): ...
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| Python SDK | pip install polos |
| TypeScript SDK | npm install polos |
| HTTP / webhook / cron / event triggers | All first-class |
| Slack integration | HITL approval channel |