The open-source Agent Development Kit for conversational agents that complete a transaction. The runtime owns the moments that must be right; the model owns the conversation. TypeScript, MCP, any model.
Forms, pickers, option chips, the confirmation summary and the create gate are drawn and read by the runtime, in code, by the whole message. A model that skips a question, invents a value, or reads "yes but change the title" as a yes cannot do so here.
The pattern
Which domain, is this a yes, which option, is it safe to write — asked as choice / noul / score questions and answered by rules first, a decision model second, an LLM with logprobs last, with calibrated confidence.
Verdict
Scripted conversations against a writes-off gateway with deterministic checks, a mock model that plays the model's part, flake classification, a gate that runs in CI with no key and no network.
Evals
Identity that fails closed, owner checks, rate limits, usage caps, pass-through metering, OpenTelemetry in the GenAI conventions, dashboards and alerts as code, graceful deploys, a widget people use on a phone.
Going to production
npx create-marg-app my-agent && cd my-agent
npm install
npm run gateway # http://localhost:8787 — the scripted mock model, no keygit clone https://github.com/marg-adk/marg.git && cd marg
npm install && cp .env.example .env
npm run gateway # http://localhost:8787Open the playground and type "ticket for Northwind: login fails on mobile". A customer picker, category chips, a details card the runtime draws and reads, priority chips, a summary with Change something, the create. Switch to a real model by editing one line in .env (LLM_CHAIN=anthropic, LLM_CHAIN=openai,anthropic, LLM_CHAIN=ollama); nothing else changes.
The numbers that justified the design, from the production agent Marg was extracted from.
| Observation | Result |
|---|---|
| Questions a model-led prompt skipped, despite the rule | 25 to 50 percent of turns |
| List items a model attested that the user never said | 1 chat in 64 |
| Typed replies a leading-word reader mis-read as a bare yes or no, blind corpus of 521 | 235 |
| The same corpus, read by the whole content | 0 |
ui blocks malformed around the fence in production traffic | about 1 in 40, all parsed |
The model must never read "yes" tells the story behind the third and fourth rows.
Use Marg when your agent must complete a transaction, not only answer; when the result is published under a real user's name; when several specialist agents sit behind one entry point; when you need to prove behaviour before and after every change without burning model credits; when you are going to run it yourself.
Look elsewhere when you want a general library for arbitrary agent graphs, your bot only answers and never writes, you want a hosted platform, or your stack must be Python end to end. Marg and other frameworks says where the line is.
Anthropic natively, and any OpenAI-compatible endpoint: OpenAI, Gemini, Vertex AI, Azure OpenAI, Bedrock, Ollama, vLLM, LiteLLM, Groq, Mistral, DeepSeek, Fireworks, Together, OpenRouter, xAI, Cohere, Hugging Face, Cerebras, Perplexity, NVIDIA, one environment line each (Models). A chain of endpoints with failover and a circuit breaker is one line, LLM_CHAIN=openai,anthropic. The scripted mock model runs everything with no key.
Marg is opinionated about one job shape, collect → validate → confirm → write, and puts the costly moments in the runtime instead of the prompt. The reader that reads "yes" by the whole message, the three-fold write gate, and the offline eval gate exist because their absence cost something in production. The pattern explains each rule and the failure it answers.
Through an MCP tool server, in any language: reads, and the triad validate_*, preview_*, create_*. The example domain's server is the reference; probe the real backend before trusting its documentation. A new domain.
Run the gate: unit tests, the OpenTelemetry check, the scaffold check, the real-write e2e, and the example suite against a writes-off gateway, all on the mock model. Every red is classified before it is fixed. Evals.
One non-root image, Compose profiles for stores, observability, decisions and the TLS edge, and a Render blueprint for a demo. Deploy, then the production checklist.