Why ScientistWorkbench
A scientific compute service shaped for agents: unattended, sandboxed, reviewed, and addressable.
Per-job sandboxes, real concurrency
Every job gets a fresh Linux sandbox (Ubuntu 24.04, up to 8 GB RAM, 10 GB disk) that is discarded afterwards. Your agents can fan out: one QC per plate, one lookup per target, one phylogeny per family — in parallel, with no queue for a shared kernel.
API / MCP native
The service is a job API first. The MCP server exposes the same jobs as tools and science:// resources, so Claude Code, Cursor, Claude Desktop and Agent-SDK agents get it with one mcp add. Same key for both.
Durable job ids and provenance
A job id resolves forever. Results carry artifacts[] with sha256, size, mime and a stable science:<job_id>/<file> ref, plus the full transcript, the execution log (every code cell with stdout/stderr) and the reviewer record. Your framework can cite what it uses.
No subscription required
Sign up with an email, get an API key, buy tokens. Compute runs on our managed sandbox platform; you never manage models, quotas or sandboxes.
Curated science environment
scipy, statsmodels, pandas, Biopython, scikit-learn, RDKit, MAFFT, FastTree, HMMER, TM-align, openpyxl, matplotlib, and skills for 24 life-science databases (UniProt, Ensembl, gnomAD, ClinVar, Open Targets, PDB, AlphaFold, GEO, cBioPortal, PubChem, ChEMBL, BindingDB, openFDA, OpenAlex, PubMed, ClinicalTrials.gov, bioRxiv, DepMap, GTEx, InterPro …). No environment setup in the prompt.
Independent reviewer
After the agent finishes, a second model re-reads the transcript and artifacts and checks each claim in the report against what actually ran. Findings come back as pass / warn / fail with evidence, so your agent knows which numbers to quarantine.
Isolation and data handling
Inputs are mounted read-only into the job's sandbox and the upload is deleted after mounting. Outputs live under per-tenant storage keys with presigned, expiring download links. Prompts and outputs are processed by our model and sandbox infrastructure providers under commercial terms that exclude training on customer data.
Unattended by design
Nobody answers questions mid-run. The prompt envelope tells the agent to make the standard, defensible choice and record it under Assumptions & caveats. That is what makes runs reproducible and what makes them safe to fire from a loop.
Where it fits
Good fits
- QC and re-analysis of CRO deliverables (plate stats, curve fits, replicate consistency)
- Structured database lookups your agent should not do from memory
- Any plot, fit, PK/PD derivation, selectivity normalisation
- Sequence and structure work: alignments, trees, ancestral states, superpositions
- Literature and evidence dossiers with cited sources
Not a fit
- Chat. Jobs are one-shot; there is no back-and-forth.
- Sub-minute latency. Lookups take minutes, QC 5–15 min, research an hour.
- GPU workloads (folding, training). A heavy tier is on the roadmap; ask us.
- Clinical decisions. See the AI disclosure.