Rasa
Conversational AI platform built on the CALM framework, with flows that orchestrate sub agents over MCP and A2A, messaging and voice channels, and self managed or managed deployment.
Rasa is a framework and platform for building conversational AI agents, text and voice assistants that a company hosts and controls itself. It began as an open source framework rooted in machine learning for understanding language and managing dialogue, and its current generation centers on an approach the company calls CALM, short for Conversational AI with Language Models, which moves away from the older method of hand labeling intents and training phrases.
The idea behind CALM is a deliberate separation of concerns. A language model handles the fluent part of a conversation, interpreting what a user means and handling things like topic changes, corrections, and clarifications, while the actual steps the agent is allowed to take are defined by the developer as Flows, structured business processes broken into clear steps. The model understands the user, but it does not invent the business logic. That design keeps execution deterministic and debuggable. The customer chooses and routes its own models, including self hosted ones, so an assistant can run with no calls to external LLMs.
The broader Rasa Platform has two parts. Rasa Pro is a pro code framework with tracing, end to end testing, PII management, and secrets handling. Rasa Studio is a no code interface for business users, with single sign on and role based access, so technical and nontechnical teams can work on the same agents. A free Developer Edition lets a team build and run one assistant within a monthly conversation limit, and paid plans are quoted by sales. Flows can hand steps to sub agents, built in ones that use tools from MCP servers or outside agents over A2A, and a Rasa assistant can itself be exposed as an A2A agent. Assistants deploy across web, mobile, messaging channels, and voice gateways.
Rasa is built to be self managed on a team's own infrastructure or private cloud, or run as a managed service, giving the customer full ownership over data, models, and behavior.
Vendor details
Canonical URL
https://rasa.com
Category
Multi-agent platform
Company status
independent
Use cases & customers
Target customers
Deployment options
In practice
You need a chatbot that can't improvise its way into a compliance problem. Rasa's CALM lets a language model interpret the user while you define the exact steps it can take as Flows, keeping execution deterministic and auditable.
Your data can't leave your infrastructure. Rasa is built to be self hosted on premises, can run with small fine tuned models and no calls to external LLMs, and gives you full ownership of models and behavior.
Your technical and business teams both need to work on the assistant. Rasa pairs a pro code framework with a no code Studio, plus a browser playground to prototype CALM agents and a free developer edition to start.
Sources & related URLs
Agentic Index coverage score
11.5 / 14 capabilities · 82%
| Integrations & Tool Calling | Full |
|---|---|
|
Flows call tools from MCP servers directly with a call step and input and output mappings, ReAct sub agents pick MCP tools at runtime, and a custom actions server built on the Rasa SDK reaches any backend API. No prebuilt connector catalog is published beyond channel connectors. Sourcerasa.com/docs/pro/build/mcp-integration.mdread 2026-09-27 |
|
| Workflow Orchestration | Full |
|
Flows define business processes as steps with conditions, links and calls to other flows, and Rasa acts as the orchestrator for sub agents inside them: built in ReAct sub agents with MCP tools and external agents over A2A, with shared context, state checks, resume after interruption and up to three retries with backoff. Sourcerasa.com/docs/reference/config/agents/overview-agents.mdread 2026-09-27 |
|
| Knowledge Grounding & RAG | Full |
|
Enterprise search indexes the customer's documents in FAISS, Milvus or Qdrant, or a custom retriever, and the enterprise search policy retrieves relevant passages when it detects a knowledge question and has the model answer from them. Sourcerasa.com/docs/pro/build/configuring-enterprise-search.mdread 2026-09-27 |
|
| Human Oversight & Guardrails | Partial |
|
A human handoff conversation pattern passes the conversation to a person, and CALM limits the model to developer defined flows so it cannot invent steps. There is no screen where a reviewer approves or rejects an agent action before it runs, and the live agent side of a handoff sits in the customer's contact center. Sourcerasa.com/docs/learn/concepts/conversation-patterns.mdread 2026-09-27 |
|
| Security, Identity & Governance | Full |
|
Studio ships SSO and role based access, set up through Keycloak, the Rasa Pro API takes token or JWT auth with role based access, and Pro adds PII management and Vault secrets. Rasa publishes no attestation. Its security page says controls are aligned with ISO 27002 and names no SOC 2 or ISO certificate, and a SOC 2 claim appears only in third party writeups, not on rasa.com. Sourcerasa.com/pricingread 2026-09-27 |
|
| Observability & Auditability | Full |
|
Tracing records the assistant's own processing, LLM calls with prompts, inputs, outputs and token usage, generated commands, flow transitions and custom actions, exported to Jaeger, an OTEL collector or Langfuse. An analytics pipeline streams conversation events through Kafka to a warehouse, and Studio adds a conversation view. Sourcerasa.com/docs/pro/improve/tracing.mdread 2026-09-27 |
|
| Memory & State Persistence | Partial |
|
Slots hold key value state the assistant collects within a conversation. Tracker stores keep every conversation, and user scoped conversations let the API list all trackers for one user, but that is the conversation record, not a memory layer with a set scope and lifetime that the assistant reads across sessions. Sourcerasa.com/docs/pro/build/assistant-memory.mdread 2026-09-27 |
|
| Deployment & Data Residency | Full |
|
The paid platform deploys self managed on premises or in a private cloud, or as a managed service, with Kubernetes deployment through Helm and deployment playbooks for AWS, Azure and GCP, and hybrid setups are also offered. The free Developer Edition runs locally or in production. Sourcerasa.com/pricingread 2026-09-27 |
|
| Prebuilt Agents, Templates & Packs | Partial |
|
Project templates such as tutorial and calm scaffold a starter assistant through rasa init, and default conversation patterns come built in, but no catalog of ready to use agents is published. Sourcerasa.com/docs/pro/tutorial.mdread 2026-09-27 |
|
| Triggers & Channel Coverage | Full |
|
An HTTP endpoint lets an outside system inject an intent into a conversation in place of a user message, after which the assistant runs its response and sends it to a named output channel, and a ReminderScheduled event lets the assistant schedule itself to act later. Most work arrives from people on messaging and voice channels such as web, Slack, Telegram, Messenger, Twilio, Genesys Cloud, AudioCodes, Jambonz and Vonage, among others. Sourcerasa.com/docs/reference/api/pro/http-api/tracker/inject-an-intent-into-a-conversation.mdread 2026-09-27 |
|
| Model Flexibility & Routing | Full |
|
The customer configures model deployments in endpoints.yml, including Azure and self hosted endpoints such as vLLM, Llama.cpp and Ollama, and a multi LLM router spreads requests by shuffle, least busy, latency, cost or usage with failover to the next deployment. Sourcerasa.com/docs/pro/deploy/llm-routing.mdread 2026-09-27 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
The Rasa Pro REST API covers status, models, conversation trackers and message injection with token or JWT auth, the Rasa SDK builds custom actions, and an A2A server exposes the assistant to outside orchestrators with an agent card at /.well-known/agent-card.json and bearer JWT auth. Sourcerasa.com/docs/reference/api/pro/rasa-pro-rest-api.mdread 2026-09-27 |
|
| Testing, Debugging & Optimization | Full |
|
End to end tests assert each step of scripted conversations with coverage reports and are meant for CI. Evals drive simulated users through YAML scenarios scored by an LLM judge plus deterministic assertions, and they are not yet suited to blocking CI. Sourcerasa.com/docs/reference/testing/evals/overview.mdread 2026-09-27 |
|
| Browser & Computer Use | Not documented |
|
The product covers channels, flows, sub agents, MCP tools and APIs, with no browser or computer use capability, and any browser work would come from an outside MCP server or A2A agent. Sourcerasa.com/docs/llms.txtread 2026-09-27 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Mantle agents can now pass context to MCP servers as metadata kept apart from the tool arguments and the model's history. Values held in skill or project memory can be sent with every tool call, and a hook can attach extra metadata to a single call. When a tool needs confirmation, the metadata is saved and sent again once the call resumes.
Bears on: MCP / tool calling / API
View sourceIn Rasa Pro 3.21.0 the copilot can plan before it builds: a prompt produces a build plan listing the agent's skills, which the user can read, refine in conversation or edit directly before approving. Only approval scaffolds and builds the agent, and plans are versioned so every refinement is kept.
Bears on: Agent capability
View sourceRasa relaunched its platform around Mantle, a new orchestrator where builders write each skill as plain language instructions and Mantle routes between skills, calls tools and tracks memory. Deterministic guarantees such as required values, confirmation steps and strict sequences are added only where they are needed.
Bears on: Workflow orchestration
View sourcePricing
Free Developer Edition · paid plans quoted by sales
Annual platform subscription, tiered by conversation volume and quoted by sales.
Included quota
The Developer Edition is a free Rasa license usable locally or in production, with one bot per company and up to 1,000 external or 100 internal conversations a month. Paid plans add Rasa Studio, SSO and RBAC, higher volume and basic or premium support at quoted prices.
What is public
The free Developer Edition and its limits are public (one bot per company, up to 1,000 external or 100 internal conversations a month, community support), along with the feature contents of Rasa Pro, the Business plan (Rasa Pro plus Rasa Studio, self managed or managed deployment, basic support) and Enterprise (full platform, premium support with 24/7 response and a customer success team). No paid price is published.
Billing mechanics
An annual platform subscription quoted by sales above a free Developer Edition. Self managed deployments carry their own infrastructure and model inference costs, and the managed service is available on paid plans.
Cost watchouts
Self managed deployments put infrastructure and, with frontier models, LLM inference costs on the customer on top of the license. The AudioCodes VoiceAI Connect IVR connector is sold separately. Paid pricing is quoted by sales.
Variable cost rationale
The platform fee is a flat annual subscription, tiered by conversation volume and quoted by sales. Self managed deployments put model inference and infrastructure costs on the customer, and those costs grow with usage. The customer controls most of that variable part. CALM can run on smaller on premises models with no external LLM calls, which holds marginal cost down, while frontier models add LLM inference costs on top of the license. The managed service is available on paid plans.
Additional watchouts
Beyond the free Developer Edition every plan is quote based. Self managed deployments put infrastructure and model costs on the customer. The free tier is capped at one bot and a low monthly conversation volume. The AudioCodes IVR connector is an additional purchase.
Overage / add-ons
The free edition is capped at one bot and 1,000 external or 100 internal conversations a month. Higher volume requires a quoted paid plan.
Sales call required
Yes, required for paid access
Free / trial
Free Developer Edition with one bot per company, up to 1,000 external or 100 internal conversations a month and community support
Lowest paid plan
Not published
Commercial notes
Rasa is built for regulated industries that want control, with self managed deployment on premises or in a private cloud, or a managed service, and CALM keeping execution to developer defined flows. Customers own the engineering, and pricing is quoted once they pass the free Developer Edition.
Key ambiguities
No paid figure is published, and the page shows a Business plan column while saying there are two subscription categories (Developer Edition and Enterprise). Self managed infrastructure and model costs are separate.
Cancellation / refund
Paid plans are an annual platform subscription quoted by sales. No cancellation or refund terms are published.
Support SLA / resale
Developer Edition gets the community forum, and Business gets basic support through a web portal, break and fix, best effort in business hours. Enterprise has premium support with 24/7/365 enhanced response times, a customer success manager and engineer, and business reviews.
Missing data
No paid price is published. Business and Enterprise are quoted by sales.
Related vendors
- CrewAI — Open source multi agent framework (Python, MIT) of role based Crews…
- Relevance AI — Platform for building and deploying a workforce of AI agents with a…
- Agent Zero — Open-source multi-agent framework that gives its agent a full Linux…
- Akka — Enterprise agentic AI platform that builds and runs multi agent…
- Artian — Enterprise multi agent AI platform for financial services, where a…
- Databricks Agent Bricks — Databricks' agent platform, now sold as Agent Bricks, for building,…
Alternatives to Rasa
The closest documented capability profiles to Rasa among multi-agent platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- CrewAI13.0 / 14Adds documented Browser & Computer Use
- Databricks Agent Bricks13.0 / 14Fuller documented coverage on Human Oversight & Guardrails and Memory & State Persistence
- Dataiku13.0 / 14Fuller documented coverage on Human Oversight & Guardrails and Memory & State Persistence
- Relevance AI13.0 / 14Fuller documented coverage on Human Oversight & Guardrails and Memory & State Persistence
- Akka12.5 / 14Fuller documented coverage on Human Oversight & Guardrails and Memory & State Persistence
- Legion Intelligence11.5 / 14Fuller documented coverage on Human Oversight & Guardrails and Prebuilt Agents, Templates & PacksRasa vs Legion Intelligence →
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded