Research you can check.

Why a workspace

A chat window forgets. A project keeps the record.

Research runs across files, tools, a cluster and weeks of sessions. A chat with plugins gives you a transcript, and you cannot check a transcript. Athar is a workspace. This is what changes.

Replay with no AI in the loop works for calls made with Athar’s packaged tools. Other runs, such as a Python script the agent writes, keep their script and outputs on record, and their receipt says Not replayable.

• 

  01“How were the affinities computed?”

  Open the Nb-C finding

  It holds the fitted rates, KD 1.248 nM, the 95% intervals and the fit residuals. It also states its weakest point: one simulated trace per concentration, so no experimental claim.

• 

  02“Can the authors reproduce them?”

  Run the saved script again

  The finding keeps the fit script and every output file. This fit ran as a plain Python script, so its receipt says Not replayable. Replay with no AI in the loop works for calls made with Athar’s packaged tools.

• 

  03“No statistical test supports p < 0.01.”

  Athar caught it while editing

  We asked Athar only to update the results paragraph. It removed the p < 0.01 claim and said why, and At a Glance lists it under Needs a closer look. It also removed a sentence about dropping Nb-D, so check its edits.

A real run in Athar on synthetic data, with the cloud model gpt-6.1-sol. The fitted KD values are within 2% of the values used to simulate the curves.

Benchmark · Terminal-Bench Science

Athar’s mean pass rate across 20 science tasks: 59%.

Across selected passing runs on 12 tasks both agents solved, Athar used 58% of OpenScience’s agent time, 34% of its tokens and 43% of its estimated cost. These totals exclude failed attempts.

These results are not a controlled ranking: trial selection, models, budgets and agent configurations differ. Resource totals exclude failed attempts; costs are estimates, not bills. Read the methods and limits →

In the workspace

Ask a question, or hand over a plan.

Two modes, four permission levels and one record. Plan the work, choose the models, watch your files, use your cluster and edit your manuscript in the same project. The examples use synthetic data and the labels from the app.

• 01 · Modes

  Choose how the agent works.

  Use Ask for questions and quick analyses in the chat. Use Pursuits for work with many steps: Athar plans first, and you approve the plan.

  Then choose how much the agent does without asking you.

  Credentials, SSH keys and shell startup files stay blocked in every mode.

  Conversation mode

  Mode	What it does	Use it for
  Ask	Discuss in this conversation. Active work continues.	Questions, quick analyses, reading
  Pursuits	Plan and track work in this conversation.	Work with many steps that you want to control

  Permissions

  Setting	What it does	Use it for
  Supervised	Asks before commands and file changes.	A new project, or data you do not know yet
  Auto-accept edits	Approves edits, asks before other actions.	Writing and editing files
  Auto	Routine actions run without asking. Python environments still ask once.	Long analyses you trust
  Full access	No approval cards for sandboxed work. Commands that need your whole account still ask.	Pipelines you know well

• 02 · Pursuits

  Approve the plan before the work starts.

  Start a Pursuit and Athar proposes a plan. For each task, set what done means, which model does the work, how the result is checked and who approves it.

  Choose when Athar checks with you: at every result, at key decisions or only at the end. Approve the plan, or ask for changes. While it runs, you can pause, stop or resume it.

  Pursuit · Proposed5 tasks

  1. Design replicated binding assays for all five nanobodies

  Done when

  A sourced proposal picks SPR or BLI, with replication, controls and QC

  Work by

  Opus

  Check the result

  Independent review of the claims

  Who approves

  Athar approves

  2. Specify the Nb-C versus Nb-A test and the replicates it needs

  Work by

  GPT-6.1

  Check the result

  Re-run the scripts

  Who approves

  I approve

  Check with me

  At key decisions

  Ask for changesApprove plan

• 03 · Models

  Several models in one project.

  Give a task one model to do the work, and a model from a different family to check it. The checker can re-run the scripts, repeat the method or review each claim against its sources.

  When the checker cannot support a claim, the claim stays Proposed and the task waits for you. In a chat, you can also switch models. The new model answers from your next message.

  Task 1 · Independent review of the claims

  Opus · work

  SPR design: 24 chips, 68.8 h of SPR time.

  GPT-6.1 · check

  C28–C30: design, replication and resource maths.

  Supported

  GPT-6.1 · check

  C31: both verification runs passed.

  Not checked

  Pursuit

  C31 stays Proposed. The task waits for you.

• 04 · Files

  Watch your files change.

  The Files panel updates while the agent works. When the agent edits a file you have open, a banner tells you. Show changes opens the diff.

  Each file keeps its versions, marked Agent or You. Compare two versions side by side, or restore an older one. One diff shows every change from this chat.

  analysis/fit_kd.py

  The agent edited this fileShow changes

  VersionsSide by side

  10:42

  After assistant turn Agent

  10:31

  Saved in editor You

  10:12

  Before assistant turn Agent

• 05 · Cluster

  Use your cluster from the same project.

  Add a host from your SSH config. Athar never asks for a key. Probe shows the host’s CPUs, GPUs, Slurm partitions, Conda and modules.

  Browse cluster files beside your local files. The agent runs commands and submits Slurm jobs only after you approve, and it logs each job it submits in the project. Ask the agent for the status of a job.

  Work you approve runs on the cluster as you, outside Athar’s sandbox. Cluster browsing is read-only and never syncs files.

  SSH hosts · hpc-loginProbe

  Slurm

  Partitions gpu, cpu, long

  GPUs

  4 × A100

  Conda

  Found

  Modules

  Found

  Remote execution approval

  sbatch fit_variants.slurm on hpc-login

  DenyAllow for this conversationAllow once

• 06 · Manuscripts

  Annotate the paper. The agent edits it.

  Open the PDF beside the chat. Select a passage, press ⌘I and write what should change. Add more annotations to the queue, then send them to the agent as one batch.

  The agent edits the manuscript source. The file’s versions show what changed.

  

  Annotations (2)

  p. 4 · “strongly enhanced”

  State the fold change and the number of replicates.

  p. 6 · Figure 3 caption

  Name the fit model.

  Review batch in composerSend to agent now

Across sessions

The next session starts from the record.

Research questions, findings and the links between them carry across sessions. A new session starts from them, not from a blank chat. Shown: the real app with the synthetic anti-PD-L1 study. Assistants, The Bench and Connections come from another sample project. Screens may change before release.

At a GlanceFieldnotesChaptersAssistantsThe BenchConnections

Pick up where you stopped.

See the latest findings, the claims that need a closer look and the project record, then resume the conversation.

Getting started

Three steps to your first checked finding.

1. Open a project folder.

   Privacy Mode is on by default. Local models run on your computer. Add a Claude or OpenAI model only if you want one.

2. Ask a question.

   Athar reads your files, runs tools and saves each finding with its sources. Open any finding to see the model, the script and the outputs behind it.

3. Continue, or hand over a plan.

   Accept a chapter when the findings hold. Start a Pursuit when you want Athar to run a plan. For each task, set Done when, Work by, Check the result and Who approves.

• Local models on your computer
• Claude models
• OpenAI models
• MCP servers, such as PubChem
• Git
• SSH hosts and Slurm clusters
• PDFs, figures and project files
• Python packages and tools

Resources

Read more.

• Perspective

  Much of life-science research cannot be traced or checked.

  Read the article →
• Our approach

  Research we can build on

  Read our approach →
• Research & perspectives

  All articles from BioVantageLab.

  See all →

Questions

Frequently asked questions.