Prismer.AI: an open source research workspace built around agents

A research workspace that spans reading, analysis and writing

Prismer.AI presents itself as an open source research platform covering the entire academic workflow, from reading papers to publishing your own, and the comparison table in the README sets the frame: it is positioned as an alternative to OpenAI Prism, with Overleaf and Google Scholar as the other columns.

The feature list is where the scope becomes concrete. Paper Reading is an AI-native PDF reader with citation graphs. The IM and Agent Protocol covers agent-to-agent messaging, groups, workspaces and real-time events. Document Parsing produces markdown output from PDFs. Data Analysis runs Jupyter notebooks with Python and R execution. Paper Writing is a LaTeX editor with real-time preview. Citation Verification checks references automatically against academic databases. And a multi-agent system orchestrates specialised agents for research work.

The argument the README makes is about coverage rather than about any single feature, and it is aimed at a specific reader. Tools like Overleaf handle writing and tools like Notion handle notes, whereas this project puts the whole chain in one place.

The project also runs live products alongside the open source part: a learning site for courses and resources, an OpenClaw deployment page that promises a one minute deployment, and a hosted cloud offering for knowledge infrastructure for agents.

Three directories carry the open source scope, and the rest of the tree is governance

The repository says explicitly that its current open source scope centres on three directories, and the tree backs that up.

The web directory holds the Next.js research workspace, described as a chat panel plus a multi-editor WindowViewer. The docker directory holds the OpenClaw-based container stack and the gateway integration. The docs directory holds the architecture, runbook, contribution and roadmap documentation, and it is where the translations of the readme live, with Chinese, Japanese, French and German versions alongside the English one.

Around those sit the files a serious project needs. There is a CITATION.cff, which is how a repository asks to be cited, a CODEOWNERS file, a SECURITY.md and a CODE_OF_CONDUCT.md, a CONTRIBUTING.md, an .editorconfig, and a CLAUDE.md, the file that tells an AI coding assistant how to work in the repository. A skills directory sits next to the workspace code.

There is also a continuous integration workflow, ci.yml, referenced by the badge at the top of the readme. The combination of a citation file and a security policy is a decent signal for a research tool that people intend to build on, even though the release history is thin.

The agent layer is a gateway with its own port and health endpoint

Self-hosting runs two containers, and the split between them is the thing to understand.

The web interface answers on http://localhost:3000 and is the Next.js frontend. The agent gateway answers on http://localhost:16888 and is described as the container gateway for LaTeX, Jupyter and OpenClaw. Aggregated service health is available at http://localhost:16888/api/v1/health.

So the browser talks to the web container, while the heavy execution happens in the gateway container, which is where the LaTeX toolchain, the notebook runtime and the OpenClaw agent core live. That separation is what lets the gateway restart without losing the user's session in the editor, and it is why the compose file is the interesting artefact rather than the frontend code.

The gateway is also the place where the agent-to-agent messaging protocol lives, with groups, workspaces and real-time events. The multi-agent system that orchestrates specialised research agents is the same machinery, and the related repositories named in the README split the work further, with an open-source agent core in Prismer-AI/luminclaw and open-source cloud harness infrastructure in Prismer-AI/PrismerCloud.

Self hosting is a clone, a .env file and one compose command

The quick start in the README is short enough to be followed exactly:

git clone https://github.com/Prismer-AI/Prismer.git
cd Prismer
cp .env.example .env

cd docker
docker compose -f docker-compose.openclaw.yml up --build -d

Two details in those commands matter. The compose file lives in the docker directory, so the working directory changes before it runs, and the file has a specific name, docker-compose.openclaw.yml, which suggests the stack has more than one variant. The build flag is present, so the images are built locally rather than pulled.

The environment file comes from the repository, and the sample file carries the structure of the whole configuration. The database is SQLite, with DATABASE_URL pointing at a file inside the container that is persisted through a Docker volume. The NextAuth secret is generated with openssl rand -base64 32, and NEXTAUTH_URL points at the local web address. The AI section names OPENAI_API_KEY, OPENAI_API_BASE_URL and AGENT_DEFAULT_MODEL, with gpt-4o as the default model.

There is a second copy of the AI settings for the browser, with the NEXT_PUBLIC_ prefix, and the comment says why: the AI-native PDF reader runs against the model from the client side. That detail has a security consequence, because a key with that prefix ends up in the browser.

One base URL covers OpenAI, Azure, vLLM, Ollama and LiteLLM

The environment example names the compatible backends directly. The comment on the AI section says that any OpenAI-compatible API works, and it lists OpenAI, Azure, vLLM, Ollama and LiteLLM among them.

That is the design decision that keeps the project self-hostable in practice. A research group with a local model server changes OPENAI_API_BASE_URL and the key, and the application talks to it the same way. A group with a hosted model changes the same two values. Anthropic support sits alongside as an optional commented setting, with ANTHROPIC_API_KEY in the sample file, for teams that want Claude models.

The rest of the configuration is layered by purpose. The required block covers the database and the authentication secret. The AI block is marked as required for chat and agent features. There is an optional data source block that switches between local data and something else depending on a flag, and the optional Anthropic entry sits next to it.

For a self-hosted deployment the practical reading is that the AI provider is a runtime decision rather than a build decision, which is why the same image works for a laptop with Ollama and for a hosted deployment with a commercial API.

The Reviewer Agent checks every reference against three databases

One feature gets its own section in the README, and it is the one with a concrete mechanism.

The problem stated is that large language models fabricate citations. The answer is a Reviewer Agent that validates every reference against academic databases, with arXiv, Semantic Scholar and CrossRef named, before it appears in your paper. So the citation check is not a spellchecker for reference strings, it is an agent that queries external indexes and asks whether the work exists.

That is a meaningful difference for anyone who has caught a hallucinated reference in a generated bibliography. It also explains why the agent protocol and the multi-agent system exist at all, since checking a reference is a task that can be handed to something with its own tools and its own context.

The caveat is on the roadmap rather than in the feature list. The Reviewer Agent appears under the in progress column of the platform table, next to the documentation site and the npm package extraction, while the paper reader, the IM API, the LaTeX editor, the Jupyter notebooks, the multi-agent system and Docker self-hosting are all in the done column.

So the citation verification is announced and designed, but the readme's own status table says it is not finished.

One archive tag, a trimmed workspace and an open licensing question

The release history is a single entry, and the release note explains it. Version v0.0.1-archive, published on 2026-03-03, is described as a pre-release archive with the code frozen for open source preparation. The last push to the repository was on 2026-06-24, and the repository is not archived.

In other words, this is a snapshot of a working product opened up, rather than a project with a release cadence. That is a normal shape for a company that has built something private and then decided to publish the source, and it explains both the quality of the documentation and the absence of versions.

The open source workspace section of the roadmap says the same thing. The workspace code extraction and trim is done, along with a dead code cleanup of roughly nine thousand five hundred removed lines, and the in-progress items are a hardcoded URL cleanup and a lite Docker image under four gigabytes.

The licensing question is the one a user should settle before anything else. The licence field on the repository describes a custom licence that GitHub cannot classify and points to LICENSE.md, while the badge in the README says MIT. Those two statements do not obviously agree, and the difference matters if you plan to build a product on the code.

Beyond that, the project is unusually open about its intentions. The readme carries a roadmap file, a runbook in the docs directory, a Discord server, an X account, a LinkedIn page and a video walkthrough titled from idea to paper, end to end.

Editorial conclusion

Prismer.AI fits a research group that wants one self-hosted environment for reading, analysing and writing papers, and that cares about citations being checked rather than generated, because the same stack carries the PDF reader, the LaTeX preview, the notebooks and the agent protocol. It is the wrong choice if you need a stable release, since the only tag is v0.0.1-archive, described as a pre-release archive with the code frozen for open source preparation, and it is the wrong choice if you need the document agent features today, because the Reviewer Agent is still marked as in progress. Before you build on it, read LICENSE.md yourself, because the licence field on the repository describes a custom licence that GitHub cannot classify while the README badge says MIT, and check the roadmap in docs before planning around the citation verification and the agent gateway.

Frequently asked questions

Official sources

1. Issues
2. Prismer-AI/Prismer on GitHub
3. Project website
4. README
5. Releases