Rasa

What it is

Rasa is a developer-first platform for building conversational AI agents. Its homepage calls it "The developer platform for enterprise AI agents" and says it "extends LLMs with your business logic to create reliable AI agents across millions of conversations", giving your team "full control to build, deploy, and manage performance across every channel." This is a chatbot builder for product teams and engineers rather than for someone who wants to publish a bot this afternoon without touching a terminal. The copyright line reads Rasa Technologies. The homepage says Rasa was named a Strong Performer in The Forrester Wave 2026 for conversational AI platforms for customer service, evaluated among 14 enterprise platforms.

What you build and how

The homepage asks "What would you like to build?" and offers a customer support agent, an internal helpdesk agent, a search agent, a voice agent, or "something else". The use cases it lists are customer experience (onboarding to loyalty), customer support (resolving routine issues and making room for deeper support work), sales enablement, and operational efficiency (automating processes and orchestrating systems). The industries list opens with finance and banking, described as "secure, compliant AI experiences trusted at scale".

One codebase, two ways in. The pitch is "One agent codebase" with the whole team involved. The Studio workspace is described this way: "Anyone can build, review, evaluate, and monitor agents visually. Changes become code in the shared repository." Developers work the other way round, in "a structured, version-controlled codebase" from their IDE, using the CLI and SDKs to build, test and debug an agent. So the visual side is real, but it writes into a repository that engineers also own. That is the main difference from a pure drag-and-drop bot builder.

Agent logic. The platform's feature list covers several layers:

• NLU: "Begin with intent-based structure, expand with AI flexibility as needed." You can start with explicit intents and add LLM behavior where it is useful.
• Agentic AI: agents that "take initiative and adapt to complexity".
• Multilingual AI: agents that adapt to language, tone and context.
• Orchestration: coordinating agents and tools "to give users the right help at the right moment".
• Voice: "Real-time voice infrastructure with enterprise-grade speed."
• MCP: a "standard way to connect with APIs as tools".

Under the hood. The homepage describes a patented dialogue management system and CALM, which it says "extends the flexibility of LLMs with the reliability of deterministic logic", plus Enterprise RAG for real-time retrieval from your trusted data and multi-agent orchestration where agents share conversational state and hand off tasks between domain skills. Its stated workflow has three steps: build the voice or chat agent in the open developer framework with a coding copilot, deploy it on your own infrastructure, then refine it in Rasa Studio and track performance across channels. A sample project on the homepage shows a banking agent wired to WhatsApp and Messenger, MCP tools, action servers and a Salesforce knowledge base.

Channels and deployment. The features section names text and messaging across web, mobile and messaging apps, and a Voice Gateway with turn-taking, timeouts and latency control; the Voice/IVR block adds barge-in and background noise handling over phone and voice channels. On hosting, the homepage says you can deploy Rasa "on your own infrastructure, private cloud, or fully offline environments", including air-gapped deployments, with control over data, security and model providers. It describes an open framework "inspired by its open-source roots" with access to prompts, policies and the underlying codebase, and built-in observability for tracing conversations and inspecting decision paths.

The Mantle starter. On the community site, the free route in is a quickstart. You request a license key, install rasa-pro, and scaffold a Mantle agent, which the page describes as "skills written as plain-language instructions, with framework-enforced guarantees where a mistake would be expensive." The quickstart is command-line work: uv sync, a Python setup check, filling in two keys in a local .env file, then rasa train and rasa inspect. A Codespaces quickstart is listed for people who would rather start in a browser-hosted environment on GitHub.

Agents and automations. The homepage also says you can "connect coding agents to the shared codebase with Rasa MCP Tools" and "automate builds, tests, and reviews". The showcase lists a Mantle starter pack of six Claude Code skills that "catch configuration mistakes before committing an agent".

Showcase builds. The community showcase holds 17 builds to inspect, described as learning examples where "each repository owns its setup and version requirements; sample tools and data need adapting before real use." Examples include a retail-banking assistant that transfers money, checks balances, manages payees and blocks cards (fictional accounts), an insurance starter pack that guides a customer through filing a claim, checking claim status and asking coverage questions, and a Docker Compose setup for running the Developer Edition as a containerized local deployment. The showcase mentions CALM flows and CALM projects in several entries.

Learning material. The community library carries guides on building an agent evaluation set, choosing an agent use case, designing a human handoff with consent, stopping customers from repeating themselves after a handoff, setting a model API key in endpoints.yml, and why text-to-speech failover returns after a voice agent restarts. The community site also links Rasa University, a Learning Centre and documentation.

Plans, claims and customers

Pricing. The homepage has a Pricing entry and "Get a demo" buttons, but the pages list no prices, so commercial terms come from a conversation with the company.

Free Developer Edition. The community site says "Rasa is a commercial platform with a free Developer Edition for individual builders." The free license allows "up to 1,000 external conversations per month, or 100 internal employee conversations, subject to the Developer Terms." You request it with an email address and agree to the Rasa Developer Terms (EULA). The Docker Compose build in the showcase lists "Docker and Rasa licence" as a prerequisite.

Recognition and customers. The homepage says "Analysts rank Rasa as a top platform for mission-critical customer service and voice AI" and cites the Forrester Wave Strong Performer placement for Q2 2026, with 5/5 scores shown for resource orchestration, app lifecycle management and pricing transparency. Customer quotes come from Providence, eUprava, Groupe IMA, Eddy Travels, Orange, nib Group, N26 and ERGO. Albert Heijn is credited with a 22% improvement in prevented contact rate and a 37% increase in quality score after migrating to CALM one flow at a time, nib Group with fallback rates dropping from 18% to 3.5% after merging five bots, and N26 with going from concept to production in four weeks. The company's aggregate claim is a 50% cost reduction in operational expenses across 50+ enterprise deployments.

Community and support. The community site lists a forum, office hours, Heroes, a showcase, a library and a Support page. A community Build Day ran on 14 September 2026 from 12:00 to 17:00 London time, with build sprints, live help, testing, demo rooms and a showcase.

Licensing. The Developer Edition is governed by the Developer Terms, and the community site calls Rasa a commercial platform.

Who it suits and who it does not

Rasa suits an engineering-led team, in a bank or insurer for example, that wants an agent it can version, test and review like software, and that needs voice and several channels under one codebase, possibly on its own infrastructure or air-gapped. The Studio workspace lets product and operations people contribute, so the work does not sit with engineers alone. A developer who wants to learn before buying has a clear path: a free license, a quickstart, 17 example builds and a guide library.

Caveats are plain. This is not a no-code tool. The quickstart is a terminal workflow with Python tooling, and the platform's own framing is "developer platform". The free license caps you at 1,000 external conversations a month, or 100 internal ones, so it is for building and testing rather than for a busy public bot. Production pricing is not published on the pages, and the showcase builds are, in the company's words, learning examples that need adapting before real use. A solo founder who wants a chatbot online by tonight is not the reader here; a team that expects to maintain agents for years may find the repository-based approach is the point.