Claude Science is Anthropic’s attempt to turn a general AI assistant into a research workbench. Announced on June 30, 2026, the beta brings literature work, code execution, scientific visualization, connectors, compute, and multi-agent review into one environment. The headline is not simply that Claude can summarize papers. It is that research outputs are intended to remain connected to the code, sources, and history that produced them.
The product arrives amid growing interest in agents that can operate scientific tools rather than only discuss science. That makes auditability central. A polished answer is not useful if a researcher cannot trace a number, inspect a transformation, or reproduce a figure.
What the beta includes
Claude Science is available in beta on macOS and Linux for eligible Pro, Max, Team, and Enterprise users. Team and Enterprise administrators may need to enable access. Anthropic describes a coordinating generalist agent with more than 60 curated skills and connectors for areas including genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.
The coordinating agent can delegate to specialist agents created by Anthropic or the user. A reviewer agent checks citations, calculations, and consistency between figures and their underlying data. This resembles a research group structure: one process plans and synthesizes, specialists execute narrow tasks, and another process critiques the result.
Claude Science can render domain-specific artifacts such as protein structures, genome-browser tracks, chemical structures, figures, and manuscripts. Anthropic says a generated figure includes the exact code and environment used to create it, a plain-language description, and message history. Researchers can request edits in natural language while retaining the code behind the revised artifact.
That is an important design choice. In ordinary chat, a chart can become an orphaned image. In an auditable workbench, it should remain a view of data and code that can be inspected.
Compute follows the research session
Scientific work often jumps among a laptop, a remote Linux machine, an institutional cluster, and on-demand GPUs. Claude Science is designed to work locally or over SSH and can submit work to existing high-performance computing resources or supported cloud compute.
Anthropic says the agent drafts a plan and asks before reaching new resources. Users can review or revoke decisions before a job is submitted. A running session keeps context in memory, so a large dataset does not necessarily need to be repeatedly loaded into the agent environment.
The company also emphasizes that large or sensitive datasets can remain on a lab’s infrastructure, with only the context needed for a step sent to Claude. That is a useful architecture, not a blanket privacy guarantee. Researchers still need to examine organizational contracts, connector behavior, logs, model data policies, and the exact content transmitted in each step.
Resource control matters as much as data control. A poorly scoped agent can submit expensive jobs, produce excessive intermediate files, or consume scarce cluster time. Labs should set quotas, require approval above thresholds, and store the full job specification beside the result.
Skills and connectors make the system domain-ready
Anthropic’s earlier Claude for Life Sciences work introduced scientific connectors and skills. Skills package instructions, scripts, and resources for a repeatable protocol; connectors expose data or tools. Claude Science combines these with an agent environment so a researcher can invoke a workflow in plain language while retaining its implementation.
This can reduce setup friction for standard tasks such as literature retrieval, single-cell quality control, molecular visualization, or querying biological databases. It can also make a lab’s validated pipeline reusable across projects.
But a skill is only as trustworthy as its code, dependencies, version, assumptions, and tests. A protocol validated on one data type may be wrong for another. Labs should pin versions, review changes, record parameters, and distinguish an exploratory skill from a validated production method.
Connectors carry a related risk: retrieved content may be incomplete, stale, licensed for limited uses, or hostile to an agent. A database result should preserve its identifier and retrieval date. Instructions embedded in a paper or web page must not override the research plan or system policy.
Multi-agent review is useful, not independent replication
Claude Science’s reviewer agent can flag untraceable numbers, suspect citations, or figures that do not match their code. Anthropic describes workflows where actor-critic pairs generate and evaluate sections of large reviews. This can catch errors earlier and force a clearer evidence trail.
However, two agents using the same model family, context, and flawed source can share the same mistake. Agent review is not independent replication. A reviewer can improve consistency without verifying that an assay was appropriate, a cohort was representative, or a causal interpretation was justified.
High-stakes conclusions still require domain experts, independent calculations, sensitivity analyses, and—where relevant—experimental replication. Researchers should label AI-generated hypotheses and never convert a fluent explanation into a clinical recommendation without the applicable validation and oversight.
What early examples suggest
Anthropic reports beta work in single-cell analysis, CRISPR screen design, protein structure prediction, cheminformatics, literature synthesis, and genetic epidemiology. One described workflow uses many custom skills and specialist agents to process thousands of papers, extract claims and quantitative findings, build an evidence database, draft sections, and generate cross-study figures.
These examples show breadth, but they are vendor-reported cases rather than controlled proof that every lab will achieve the same speed or quality. Adoption decisions should use a representative internal benchmark: known papers, known calculations, deliberate citation traps, ambiguous data, and a task where the correct answer is “insufficient evidence.”
Useful evaluation measures include source recall, citation precision, numerical reproducibility, figure-to-code consistency, rate of unsupported claims, compute cost, expert review time, and how often an agent correctly escalates uncertainty.
The evaluation set should include negative results and retractions, not just clean success cases. Ask the system to distinguish a preprint from peer-reviewed evidence, detect when two papers reuse the same cohort, and preserve units through a calculation. Run the same task after a skill or model update and compare the artifact, citations, and environment. If a result changes, the audit trail should make the cause discoverable rather than presenting the newest output as timeless truth.
Publication workflows also need authorship rules. A lab should state where AI assisted with search, code, figures, or drafting according to the relevant journal and institutional policies. Claude’s activity history can support disclosure, but investigators must decide what contribution is scientifically material and remain accountable for the submitted work.
A defensible research workflow
A practical process separates exploration from record:
- Define the research question, inclusion criteria, and prohibited conclusions.
- Register sources with stable identifiers and retrieval dates.
- Version the agent instructions, skills, code, environment, and parameters.
- Keep raw data immutable and write derived outputs to a new location.
- Require reviewer checks, then independent human review.
- Export a reproducibility package containing code, logs, citations, and environment details.
- Label unresolved uncertainty in the manuscript and any public derivative.
Claude Science can reduce the labor of maintaining this chain, but the institution remains responsible for it.
From an approved research artifact to Medux outputs
Once a manuscript or report has passed scientific review, a separate communications workflow may convert it into accessible formats. Claude could prepare a plain-language script, while Medux performs document and media operations through an independently configured MCP connection. This downstream use should never feed stylistic edits back into the scientific record without review.
The Medux tutorial for converting DOCX to PDF with Claude MCP provides one publication step. The PDF merge and split workflow can assemble or extract approved sections. For an audio summary, the Codex MCP voice cloning and TTS tutorial shows a separate synthesis process, which should use an authorized voice and a script checked against the paper.
There is no claimed native integration between Claude Science and Medux. The safe handoff is explicit: freeze the approved source, generate a communications copy, record each conversion task, verify the output, and clearly distinguish a narrated summary from the original research. Used this way, media tooling extends access without weakening the audit trail Claude Science is designed to build.
Keep the approved manuscript version, communications script, conversion task IDs, and reviewer sign-off together in a release record. If the science changes, regenerate derivatives from the new approved source rather than patching an old narration informally.