01 / Persistent identity
A habitat is a place, not a wrapper.
Bundle persona, config, memory, skills, secrets, tools, and sub-agents into one directory that can move across interfaces without losing itself.
Outline-first technical modernism / research toolchain
Every model lives inside a different world.
Umwelten gives you agent environments that can remember, observe, test, and compare themselves — across models, providers, sessions, and interfaces.
Habitats
Persistent environments with persona, tools, memory, sessions, and sub-agents.
Evals
Run the same task across models, score results, and compare cost and timing.
Sessions
Index prior work, surface learnings, and feed them back into future runs.
Why Umwelten
The point is not just a model. It is the world around the model.
Most AI surfaces flatten everything into a chat box. Umwelten treats an agent as a living runtime with identity, tools, memory, and measurable behavior.
02 / Comparative intelligence
Run the same task everywhere.
Evaluate one prompt across multiple providers and local models, then compare output quality, cost, timing, and failure modes instead of guessing.
03 / Reflective memory
Turn sessions into feedback.
Index Claude Code or Cursor history, search it semantically, and extract useful patterns that make future environments sharper.
04 / Interface portability
One environment, many doors.
Use the same agent runtime in the CLI, on Telegram, in Discord, or on the web without rebuilding the world each time.
Runtime map
From prompt surface to inspectable agent stack.
The architecture section leans diagrammatic rather than decorative: each layer exists so the agent can observe, remember, or compare itself better.
Layered system
Interfaces
CLI, Telegram, Discord, and web all connect to the same habitat instead of spawning isolated personalities.
Habitat core
Persona, loaded skills, tools, memories, sessions, and managed sub-agents live together in one inspectable environment.
Eval engine
Prompt suites, scoring, caching, resuming, leaderboards, and cost/timing metrics make model comparison routine.
Session analysis
Claude Code JSONL and Cursor SQLite histories become searchable evidence instead of dead logs.
Providers + locals
Hosted frontier models and local Ollama-style models can be tested side-by-side with the same evaluation contract.
Typical flow
How it behaves in practice
01
Create a habitat
Initialize an agent environment with the tools, sessions, memory, and persona you actually want to persist.
02
Run the same prompt across models
Use evals to expose differences in reasoning, instruction following, latency, and cost before choosing a default stack.
03
Analyze past sessions
Pull previous work into the loop so the environment can remember what worked, what failed, and where behavior drifts.
04
Reuse everywhere
Keep the environment itself stable while the surfaces around it change.
Install + examples
The commands should feel close enough to use immediately.
This section keeps the product page grounded in real operator behavior instead of abstract promise language.
Try it now
Three entry points: create a habitat, run a cross-model eval, compare local and cloud.
npx umwelten habitat npx umwelten eval run \ --prompt "Explain why the sky is blue in exactly three sentences" \ --models "google:gemini-3-flash-preview,openrouter:openai/gpt-5.4-nano,openrouter:anthropic/claude-sonnet-4.6" \ --id "sky-test" --concurrent npx umwelten eval run \ --prompt "Write a haiku about recursion" \ --models "ollama:qwen3:30b-a3b,openrouter:openai/gpt-5.4" \ --id "local-vs-cloud" --concurrent
Framing line
Not just another chat wrapper.
Build agents that can keep a world, inspect that world, and learn from what happened inside it.Use this mockup tone as: technical, warm, assertive, and a little poster-like.