NewInteractive Agents Live

Interactive CLI & SDK

Operate the entire platform. As code.

The iai CLI and the Python SDK, designed to build and run Interactive Agents.

kyc-verifier · iai

Created from your codebase.

JSONYAML
01Control
02Agents
03Context
04Observe
05Infra

Control the entire platform.
From your terminal, or your coding agent.

The iai CLI and the Python SDK give you programmatic control of the entire platform: your agents, their context, and the infrastructure they run on. Drive them yourself, or point your favourite coding agent at them.

Control

Control the whole platform.
As code.

The iai CLI and the Python SDK are one programmatic interface to the entire platform and every agent, built for human developers and AI agents alike. Point your favourite coding agent at it, like Claude Code, or build with Co-Pilot.

One interface

The CLI and SDK control the entire platform and your agents.

Humans and agents

The same surface for developers and for AI coding agents.

Coding-agent native

Point Claude Code or any MCP agent at the platform.

Or use Co-Pilot

Drive the same objects in conversation.

claude-code
# point your coding agent at the platform
$ iai mcp add interactiveai
 
> build the eu-refund agent and ship it to staging
 
iai agents create --file eu-refund.yaml
iai stacks sync --env staging
 
✓ driven by your coding agent · same CLI, same SDK

Agents

Build any agent.
As a declarative manifest.

Declare the agent's behavioural logic, routines, policies, prompts, glossaries, macros, in one agent config manifest. No per-agent code. Create and manage agents with a single iai command.

iai agents

Create, update, and manage agents from the CLI.

Declarative manifest

Create the agent from one declarative YAML manifest.

Routines & policies

Multi-step processes and single-step rules, declared inline.

Prompt repository

Version-controlled prompts with prod, staging, and dev labels.

eu-refund.yaml
name: EU Refund Agent
id: eu-refund
version: 0.0.1
config:
llms: { default: anthropic/claude-haiku-4.5 }
context:
system_prompt: prompts/eu-refund.md
policies: [eu-refund-compliance]
routines: [refund-intake]
 
$ iai agents create --file eu-refund.yaml
 
✓ created · eu-refund · v0.0.1 · policy + routine bound

Context

Manage context.
As versioned components.

Policies, routines, glossaries, prompts, every context component managed as a versioned, schema-enforced object. The schema is checkable by developers and interpretable by AI agents, so generated context can never be malformed.

Versioned components

Every edit is a new version, reversible at any point.

Schema-enforced

Validated on create, so context can't be malformed.

Agent-readable

The same schema AI agents author against.

Policies & routines

iai policies, routines, prompts, glossaries, macros.

routines/refund-intake.yaml
routine: refund-intake
version: 4 # every edit is a new version
schema: validated on create
 
$ iai routines create --file routines/refund-intake.yaml
$ iai policies list
 
✓ schema valid · v4 · versioned, reversible, agent-readable

Observe

Inspect any run.
Human or agent.

iai traces returns the full hierarchical trace: every observation, model route, score, and cost on one record. Let other AI agents inspect running agents, and create autonomous evals that score them continuously.

Traces

Hierarchical traces with full attribution per turn.

Human or agent

Developers and other AI agents read the same record.

Autonomous evals

Score running agents continuously against your bar.

Scores & metrics

Read and write scores; trace-derived metrics.

~/eu-refund
$ iai traces get trace_8f2a
 
contacts.get 120ms
rule check
escalated to compliance
 
✓ 1 turn · 6 observations · 1 trace · audited

Infra

Deploy the infrastructure.
Without managing it.

Define services and databases in a stack config, then deploy them with one command, or manually from the terminal. The power of Kubernetes under the hood, behind a CLI, so you focus on building agents that deliver value, not on infrastructure.

Services & databases

Deploy and manage both from one place.

Kubernetes, hidden

Autoscaling and replicas, without the operational overhead.

Stack config or terminal

Sync from a stack config, or deploy by hand.

Managed Postgres

Postgres with pgvector for retrieval.

stack.yaml
services:
- name: eu-refund
replicas: 3
autoscale: true
databases:
- name: refunds-db # managed Postgres + pgvector
 
$ iai stacks sync
 
✓ services + databases live · 3/3 replicas healthy
Most platforms force a choice: the chat for your operators, or the SDK for your engineers. Here they're the same primitives, composed two ways. Co-Pilot drafts in conversation; your engineers compose in code. Same audit trail. Same approval pipeline. Same shipped artefact.
Pedro Mota Carmo

Pedro Mota Carmo

Enterprise Implementation Lead, InteractiveAI

Three surfaces · One platform

Operated by the engineers who own the runtime.

No new mental model. The iai CLI and your CI compose the same objects Co-Pilot composes in chat; the Python SDK traces and integrates the running agent. One versioned catalog, one trace.

Python SDK

Instrument and integrate in Python.

  • pip install interactiveai
  • Trace your own code on OpenTelemetry, with full spans and attributes.
  • Integrate channels with a running agent: open sessions, stream events, set customer variables.
  • Native LangChain & LangGraph, plus LlamaIndex, CrewAI, AutoGen, OpenAI Agents, Pydantic AI, and the Claude Agent SDK.

iai CLI

Build and operate from your terminal.

  • iai agents create: the agent from a versioned manifest.
  • iai routines · policies · prompts · macros · glossaries: behavioural config as code.
  • iai stacks sync: services and databases, from a stack config or the terminal.
  • iai datasets · dataset-runs · scores · traces: evaluate and audit every run.

Coding agents

Point your coding agent at the platform.

  • Expose the platform to Claude Code, or any MCP-capable agent.
  • Have it create, deploy, and monitor your Interactive Agents.
  • Every component is schema-enforced, so generated context can't be malformed.
  • Same CLI, same SDK, same governance as a human developer.

One versioned catalog · One trace · Composable in chat or in code

CLI & SDK, answered.

The questions engineering leads and platform teams bring before building on a new AI runtime.

No. Both compose the same primitives: policies, routines, prompts, stacks. Co-Pilot drafts them in conversation; the iai CLI applies them from a versioned manifest, and the Python SDK instruments and integrates the running agent. Same versioned catalog, same approval pipeline, same trace.

Start Building

Build your first agent from the command line.

Author policies, ship through iai stacks sync, and stream traces to your stack, all from your dev loop. Read the docs or start building now.