ScientistWorkbench — product guide

ScientistWorkbench runs scientific analyses for your agents. You submit a prompt plus input files (assay exports, sequences, tables), a sandboxed science agent plans and executes the work with real code, and you get back a markdown report, figures, tables, scripts, an execution log and an independent reviewer record. Every job has a durable id you can come back to.

Access

• Account: accounts are invite-only during the private beta — request access, and the invitation email carries a link that opens signup for your address. Every account is a separate tenant with its own API keys, balance and job history. Nothing is shared between tenants.
• API keys: issued from the dashboard, shown once, stored hashed. Send them as Authorization: Bearer wb_….
• Two interfaces over the same service: the HTTP job API (https://api.scientistworkbench.com, see the API guide) and an MCP server at https://api.scientistworkbench.com/mcp for agent frameworks and coding assistants.

Tokens

• Prepaid tokens, bought in packs or granted as a trial. A job debits tokens in proportion to the model, tool and sandbox usage it consumed, plus a platform fee for jobs that complete on their own. There are no per-job caps: a running job is never stopped for budget; its usage is debited when it ends. A balance may dip below zero and then blocks the next submission until topped up.
• Typical usage measured on the production engine: CRO assay QC ≈ 185k tokens (≈ 11 min); research-grade phylogeny ≈ 1.1M tokens (≈ 39 min). The billing block on every result shows tokens_debited and balance_after_tokens.

Job classes

class	model tier	wall-clock backstop	intended for
specialist (default)	standard	60 min	assay QC, curve fits, plate statistics, PK parameters, plots over supplied files
lookup	standard	30 min	database and literature questions (DepMap, Open Targets, ChEMBL, UniProt, PubMed, trials)
research	frontier	180 min	open questions needing planning, many tools and synthesis

Classes set the model tier and a wall-clock backstop only.

What you get back

GET /jobs/{id}/result returns summary_md (Objective · Method · Results · Assumptions & caveats · Files), artifacts[] (filename, mime, sha256, size, science:<job>/<file> reference), reviewer_findings[] (an independent pass that checks the report's claims against the execution record; verdict pass / warn / fail with evidence), usage, billing, warnings, disclosure. Artifacts download from GET /jobs/{id}/artifacts/{file} (a short-lived signed link). _execution_log.json is the derived record of every command the agent ran.

Limits

• 2 jobs in flight per account by default (raise on request); 60 requests/min and 10 submits/min per key.
• Inputs: inline base64 ≤ 20 MB, https URL fetch ≤ 500 MB, multipart upload ≤ 600 MB.
• Sandbox: Linux, ≤ 8 GB RAM, ≤ 10 GB disk, no GPU (a heavy tier is on the roadmap).
• Artifacts are retained 90 days; the result document (summary, findings, billing) is kept while your account exists.

Data handling

• Inputs are mounted read-only into a per-job sandbox and the upload is deleted after mounting. Sandboxes are destroyed when the job ends; nothing carries over between jobs or tenants.
• Outputs live under per-tenant storage keys; downloads are presigned and expire.
• Your prompts, files and outputs are confidential: never used to train any model (ours or a provider's), never disclosed or used in examples without your written consent, never shared between customers, and deletable on request. The full commitment is section 3 of the privacy policy; a DPA is available.

Disclosure

AI-generated analysis. Outputs may contain errors; verify numbers against the execution log and source files before relying on them. Not a substitute for review by a qualified scientist, and not medical advice: any use affecting patient care or healthcare decisions requires review by a qualified professional.