Kosmos is an open-source project that implements an AI agent for autonomous scientific discovery. It rebuilds, as working code, the architecture described in the paper “Kosmos: An AI Scientist for Autonomous Discovery” published on arXiv in November 2025. The re-implementation was created by independent developer jimmc414 and is published on GitHub under the MIT license. The goal is to hand the cycle a human researcher repeats ─ forming hypotheses, reading literature, writing and running experimental code, validating results ─ to a group of agents. The research process is recorded in a knowledge graph, so you can trace afterwards which finding came from which evidence.
One point deserves attention: this is not an official implementation by the paper’s authors or by Edison Scientific (the company behind the original work), but a community re-implementation. The developer explicitly states that the accuracy and the discoveries reported in the paper have not yet been reproduced. As of August 2026 the version is 0.2.0-alpha, so it is best treated as experimental research software.
Key Features
- Research agents with divided roles: Separate agents handle research direction, hypothesis generation, experiment design, data analysis, and literature analysis, and their outputs are combined as the research cycle runs
- Integrated literature search: Searches across ArXiv, PubMed, and Semantic Scholar to support hypotheses and check prior work. It fits well with fields that have a deep body of published papers, such as biology and materials science
- Code execution in a Docker sandbox: Analysis code written by the agents runs in an isolated environment with limits on CPU, memory, and execution time, so runaway code does not affect the host
- Knowledge graph with Neo4j: Relationships between concepts and the discoveries derived from them accumulate as a graph. Neo4j is optional; without it, artifacts are left as JSON files
- Hierarchical context compression: Context that grows during long research cycles is compressed through three tiers of summarization, allowing longer exploration within a limited context window
- Switchable LLM providers and cost caps: Anthropic Claude is the default, and you can switch to OpenAI or to local models via LiteLLM (such as Ollama). A budget ceiling can be set before a run, and processing stops when it is about to be exceeded
Pricing
| Plan | Price | Key points |
|---|---|---|
| Open-source edition (MIT license) | $0 | Pull it from GitHub and run it in your own environment. There is no fee for the software itself |
| External costs at runtime | Usage-based | Usage fees for the LLM APIs you call (Anthropic, OpenAI, and so on) apply separately. Using local models avoids API charges |
Pricing is as of August 2026. Please check the official repository for the latest information.
Note that Edison Scientific offers a commercial service also named “Kosmos”. It is a separate product from this project, and its pricing is not published ─ you need to contact them.
Pros & Cons
✅ Pros
- Lets you read the design of an autonomous research agent as working code rather than as a paper description
- The MIT license leaves you free to modify it and to embed it in your own research workflow
- A Docker sandbox and a budget ceiling are built in from the start, which limits the risk of running it unguarded
- Switching to local models lets you evaluate it without sending data to external APIs
- Literature search, code execution, and knowledge graph are assembled into a single pipeline
⚠️ Cons
- It is an alpha release, so specification changes and bugs should be expected
- The developer himself states that the accuracy and the discoveries reported in the paper have not been reproduced
- Getting it running is heavy: Python 3.11 or later, Docker, and Neo4j if needed
- It calls LLM APIs heavily, so the cost of long runs is hard to predict
- Any “discovery” it outputs must be verified by a human and cannot be used as a result on its own
- It is an individually maintained repository, so long-term maintenance is not guaranteed
Comparison with Similar Services
| Item | Kosmos (this project) | Kosmos (Edison Scientific) | The AI Scientist (Sakana AI) | AI co-scientist (Google) |
|---|---|---|---|---|
| Delivery | Open source (self-hosted) | Commercial service | Open source | Program for researchers |
| Main audience | Individual developers and researchers | Corporate R&D teams | Automating AI research | Biomedical researchers |
| Code execution | Docker sandbox | Runs on the platform | Local execution | Not disclosed |
| Record of knowledge | Neo4j knowledge graph | Research reports | Paper-style output | Hypotheses and proposals |
| Price | Free (API costs separate) | Contact required | Free (API costs separate) | Contact required |
Who Is It For
- Developers who want to understand at the code level how an autonomous research agent is assembled
- People who want to try a pipeline that connects paper search, data analysis, and hypothesis generation on their own research topic
- Anyone looking for implementation examples of multi-agent design and context compression
- Research teams that want to evaluate an autonomous agent with local models, without handing data to an external service
- People who want to check where AI-driven scientific discovery actually stands, through an implementation rather than marketing copy
Conversely, it is not a fit if you want a finished, ready-to-use research assistant, or if you cannot spend time on environment setup.
Summary
Kosmos turns a much-discussed paper on autonomous AI scientists into code that anyone can run locally. Literature search, sandboxed execution, a knowledge graph, and context compression are all present, which makes it valuable as material for studying how such agents are designed. At the same time, keep in mind that it is an alpha release and that the developer states the paper’s results have not been reproduced. The right distance is to treat it as a testbed, not as a finished product that will do research on your behalf.