Boltz API

Three of the ten tools do. The boltz_start_structure_and_binding, boltz_start_small_molecule_screen and boltz_start_protein_screen tools submit GPU jobs, and Boltz bills by compute measured at run time. The three matching estimate tools return a price without creating a run or using any GPU, so an estimate call is free and a start call is not.

Five scopes, four of which are identity claims. Boltz advertised openid, profile, email, offline_access and compute:run on 2026-08-22. Only compute:run is an application scope, and it is the single grant that covers both reading a job's status and submitting a new billable one. No scope separates the read tools from the spending ones.

Set a workspace spending limit, not an MCP setting. Boltz documents a hard lifetime cap per workspace, denominated in milli-USD, where 1000 means USD 1.00. Boltz reserves each job's estimated cost against the ledger before the work starts and rejects anything over the cap with the error workspace_spending_limit_reached. Setting the limit to zero blocks new billable work.

Boltz's pricing page lists USD 0.025 per small molecule for the small-molecule pipeline, and a crop-size ladder for protein design and Sandbox running from USD 0.025 at 0-256 residues to USD 0.40 above 2048. The Professional plan carries no subscription and includes up to 200 free predictions each month. A library screen multiplies the per-unit figure by the library size.

Not with the ten tools Anthropic lists. Boltz's REST API has an irreversible delete-data endpoint on every prediction, design run and library screen, and no tool among the ten carries a delete verb or maps to it. Data still disappears on its own schedule: workspace retention defaults to seven days after a run reaches a terminal state, capped at fourteen.

No, they are two different servers with two different tool sets. The remote server at mcp.boltz.bio is what Anthropic's directory lists. Boltz also ships a local .mcpb desktop extension whose eleven tools share exactly one name with the remote listing, boltz_get_guidance, and which shells out to a boltz-api CLI you install yourself.

Yes, with a test-mode API key, though not over the MCP connector. Boltz documents keys prefixed sk_bc_admin_test_ and sk_bc_ws_test_ that run the full request, poll and fetch lifecycle against real endpoints and return synthetic results with no GPU usage and no billing. The MCP server signs in with OAuth rather than an API key, so test mode is a CLI and SDK path.

One request limit and one queue cap. Boltz documents 100 requests per 10-second sliding window scoped per organization, returning HTTP 429 with Retry-After. Separately, queue caps reject new submissions at 5,000 pending predictions and 500 pending pipeline runs per organization, or 1,000 and 100 per workspace. Test-mode submissions are exempt from the queue caps.