Inworld AI vs Thread AI (2026)

Customers can add their own tools. A Realtime API session registers function definitions in session.tools with a tool_choice setting, and the agent calls them mid conversation to fetch live data or trigger actions the customer's code carries out. Tools can change mid session.

SourceInworld, docs.inworld.ai (Using Realtime Models: Function calling)read 2026-09-27

Functions make authenticated API calls over REST or gRPC with OAuth or API key schemes and versioned input and output schemas. They are kept in a Function Registry where they can be tested with sample data, alongside predefined Functions that need no setup. Connections pull records from databases and APIs such as Salesforce, Google Drive or Shopify using stored Secrets.

SourceThread AI, threadai.com/docs (Lemma Core Concepts)read 2026-09-21

Inworld offers a fixed voice pipeline plus request routing, with no workflow model. The Realtime API runs speech to text, the chosen LLM and text to speech as one loop Inworld operates, and the Router picks a model per request by conditional rules, splits and failover. Neither sequences multi step agent work that the customer designs. The graph based Agent Runtime Inworld offered earlier is no longer offered, and its old address, inworld.ai/runtime, now redirects to /realtime-api.

SourceInworld, docs.inworld.ai (Realtime Overview, Router Core Concepts)read 2026-09-27

A Worker is a versioned, immutable sequence of States. States include Call, Switch for branching, Parallel, Do/While loops, ForEach in parallel or in sequence, Do/Listen and Handoff, each with a retry policy, timeouts and fallback Functions, and they combine deterministic Functions with LLM Functions in cyclic or acyclic workflows. The platform itself carries the branching, retries and routing across steps.

SourceThread AI, threadai.com/docs (Lemma Core Concepts)read 2026-09-21

People reach the agent over voice and the web. Realtime sessions open over WebSocket or browser WebRTC, and phone calls reach the agent through Twilio Media Streams or ConversationRelay. Every session starts when a user connects or calls or the customer's application opens it. No schedule, event or webhook starts an agent.

SourceInworld, docs.inworld.ai (Twilio Integration, WebRTC Connection)read 2026-09-27

A Connection pulls data from a database or API on a preconfigured schedule or as a stream and starts a Run of its Worker whenever new data arrives, a Do/Listen State waits for incoming signals, and every published Worker also runs from its own endpoint. Streaming, batch and interactive execution are supported, so work reaches the agent without a person initiating it.

SourceThread AI, threadai.com/docs and platformread 2026-09-21

There is no retrieval over the customer's content. Inworld covers TTS, STT, the Realtime API and the Router. The Router's web search grounds answers in the public web, which is not the customer's corpus, and context otherwise arrives in the prompt or through the customer's function calls. There is no knowledge base, index or ingestion.

SourceInworld, docs.inworld.ai (llms.txt, Router Web search)read 2026-09-27

Connections sync data from sources such as Google Drive, Salesforce or Shopify into Workers, including a Worker that generates text embeddings, and Functions call the customer's own services, so the customer can assemble retrieval inside a workflow. Lemma itself maintains no retrieval index or store over the customer's knowledge.

SourceThread AI, threadai.com/docs and platformread 2026-09-21

Memory lasts for the session, with a stated lifetime, and the long term store is the customer's. With providerData.memory on, the server extracts durable facts and a running summary during a session and holds them for the life of the WebSocket plus a 15 minute idle TTL. After that, server side memory is gone, and long term memory means the customer saves the facts in its own database keyed by user and seeds the next session. What Inworld itself keeps is conversation state inside a session.

SourceInworld, docs.inworld.ai (Long-term Memory)read 2026-09-27

Each Run keeps its own Context, where the inputs and outputs of every State are stored and shared across that Run's steps, and a context layer is embedded in and across workflows. No memory layer with a stated scope and lifetime beyond a single Run is described.

SourceThread AI, threadai.com/docs and platformread 2026-09-21

The customer places guardrail building blocks, but there is no approval step. Router moderation endpoints classify standalone text or whole conversations against OpenAI compatible and AILuminate safety categories, so the customer can screen user input before the LLM and filter output before users see it, and tool_choice limits which tools a session may call. Nothing holds an agent action for a person, and the constraints are ones the customer sets.

SourceInworld, docs.inworld.ai (Moderations, Using Realtime Models)read 2026-09-27

A Handoff State pauses a Run for human intervention at any point the builder places it, with preconfigured timeouts, and sends a Notification by SMS, WhatsApp, email or webhook. The person's reply resumes the Run with their input, and the builder chooses whether a reply from an outside channel can unpause it or whether sign-off must happen in the Lemma console. Lemma also has a review queue and approve, reject or edit worker functions, so the customer decides which steps need sign-off, who signs off and what happens on timeout.

SourceThread AI, threadai.com/docs and homepageread 2026-09-21

Access runs through team accounts with organization roles managed by Owners and Admins, SAML 2.0 SSO through the customer's identity provider on verified domains with an option to require SSO, and SCIM 2.0 provisioning of members and roles on Enterprise. Inworld states it is SOC 2 Type II certified and HIPAA and GDPR compliant, and SOC 2 Type II applies across deployments, including the regions. Zero data retention can be set per organization or workspace.

SourceInworld, docs.inworld.ai (SAML Single Sign-On, Team Management, Regions) and inworld.ai/securityread 2026-09-27

Access to agents and resources is governed by integration with the customer's identity provider using its existing groups and permissions, with a management layer for non-human identity permissions and PII redaction at every level. Thread AI states SOC 2 Type 2, HIPAA, GDPR and CJIS compliance and runs a trust center.

SourceThread AI, threadai.com homepage and platformread 2026-09-21

Reporting covers usage and routing, with no trace of each session. The Portal's Usage page breaks spend down by model across TTS, STT and LLMs, and Router data integrations export routing exposure to BigQuery or Mixpanel. There is no view of a session's turns, tool calls and outputs, so reporting stays at the aggregate level.

SourceInworld, docs.inworld.ai (Usage, Router Data Integrations)read 2026-09-27

Every agent decision is logged and traceable, every Worker is versioned and traced from its first deployment, and each Run's Context stores the inputs and outputs of every State, so system access, inputs and AI reasoning can be inspected run by run.

SourceThread AI, threadai.com platform, homepage and docsread 2026-09-21

Customers can choose named regions. Beside the US deployment, Inworld runs fully isolated EU and India regions, each a complete stack with inference, APIs, Portal, databases and storage in the region, provisioned per organization on Enterprise. VPC deployment is also available on Enterprise.

SourceInworld, docs.inworld.ai (Regions) and inworld.ai/llms.txtread 2026-09-27

Lemma deploys in the customer's virtual private cloud, on premises or as multi-tenant SaaS, and it is agnostic to cloud provider across any topology or network.

SourceThread AI, threadai.com homepage and platformread 2026-09-21

There are no packaged agents. Inworld offers voices, through a shared Voice Library, voice cloning and voice design, and quickstarts, not agents a customer adopts. The templates that came with the earlier Agent Runtime are no longer offered.

SourceInworld, docs.inworld.ai (llms.txt, Voice Sharing, Realtime Quickstart)read 2026-09-27

Templates are pre-defined Worker configurations for recurring patterns, such as agentic loops built from Do/While and ForEach States, available for use and customization with availability set per stack, and predefined Functions and shared building blocks can be reused across teams. These are patterns and components the customer assembles into its own Workers, and there is no set of complete ready made workers to adopt.

SourceThread AI, threadai.com/docs and homepageread 2026-09-21

The customer chooses and routes models. The Router exposes more than 220 LLMs from OpenAI, Anthropic, Google, Meta, Mistral and others, plus open models hosted by Inworld, behind one OpenAI compatible endpoint, with per request model choice, conditional routing on request metadata, traffic splitting and provider failover the customer configures. A Realtime session sets its LLM by modelId and can switch mid session.

SourceInworld, docs.inworld.ai (Router Introduction, Traffic Splitting, Using Realtime Models)read 2026-09-27

Builders choose the model for each AI Function, from predefined Functions covering current models and models with compliance guarantees to any model API they register with their own credentials. Fallback Functions switch to a backup AI provider when the primary fails, and Thread-managed credentials are configurable per tenant.

SourceThread AI, threadai.com/docs and platformread 2026-09-21

An API and SDKs cover Inworld's own platform. REST and WebSocket APIs serve TTS, STT and the Realtime API, which also runs over WebRTC. The Router is compatible with OpenAI and Anthropic and has a management API to create, list, update and delete routers, and there are Python and Node.js SDKs and an Inworld CLI with a built in MCP server that exposes its operations as tools.

SourceInworld, docs.inworld.ai (llms.txt, API reference, Inworld CLI)read 2026-09-27

Lemma offers REST, gRPC, GraphQL and SFTP APIs, SDKs and embeddable white-label components, builders work through a code-first SDK or the APIs as well as the visual builder, and every published Worker is executable from its own endpoint. The full API reference sits behind the customer console.

SourceThread AI, threadai.com platform, homepage and docsread 2026-09-21

Customers can optimize the deployed agent in a loop. Router traffic splitting sends weighted shares of live traffic to model variants for A/B tests, and custom metrics built from the customer's imported product events, from BigQuery or Mixpanel, are attached to routers to measure each variant's impact. The variable under test is the agent's model, and the outcome measured after deployment is the customer's own KPI.

SourceInworld, docs.inworld.ai (Traffic Splitting, Custom Metrics)read 2026-09-27

The Function Registry tests a Function against sample input before it is added to a Worker, and teams can experiment, test and iterate before deploying to production. There are no scored test cases, evaluation harness or quality gate returning a verdict on agent behavior.

SourceThread AI, threadai.com/docs and platformread 2026-09-21

There is no browser, desktop or computer control. The browser appears only as a WebRTC client for voice sessions.

SourceInworld, docs.inworld.ai (llms.txt, WebRTC Connection)read 2026-09-27

Lemma reaches systems through API Functions, Connections and SFTP, with no browser, desktop or computer control by an agent.

SourceThread AI, threadai.com platform and docsread 2026-09-21