Inferable

• CollectionDurable Execution & Scheduling
• Classificationagent-native
• MCPUnavailable
• Agent SkillUnavailable
• InterfacesREST · CLI · SDK · API

• Website ↗
• Official repo ↗
• Full dossier on GitHub ↗

What it does

Inferable is a managed durable-execution runtime built specifically for AI workflows and agents. It provides three tightly integrated primitives that, taken together, distinguish it from generic workflow engines:

1. Durable execution — every step is checkpointed; agent crashes resume from the last successful step instead of re-running expensive LLM calls.
2. First-class HITL — await human_approval(...) suspends the workflow without blocking a process; humans approve via Slack or email; the workflow resumes when they reply.
3. Structured outputs + versioned workflows — outputs are schema-validated; workflows are versioned so deployments don’t break in-flight runs.

It overlaps in spirit with Trigger.dev / Inngest / Restate (already in this catalog) but takes a more opinionated agent-shaped stance: HITL approval is built in (not a pattern), structured outputs are first-class, and the SDK assumes LLM/tool calls are the work.

Agent-facing interfaces

• Catalog protocols: REST, CLI, SDK, API
• MCP: ⚠️ Not yet published as a first-party MCP server. The SDK and REST API are the primary surfaces for now; community MCP wrappers exist.
• Agent Skill: ⚠️ Not yet published.
• Onboarding type: cli

How agents onboard

# Self-hosted (Docker Compose) or hosted SaaS — pick one.
# SDK install:
npm install @inferable/sdk
# or
pip install inferable

# Then wrap a function as a durable workflow step:
from inferable import workflow

@workflow.step()
def call_tool(...): ...

See the full onboarding section in the dossier.

Protocol surface