Boltz built its drug-discovery API ‘for agents as much as for people’

Since AlphaFold 2 cracked protein-structure prediction in 2020 and 2021, a growing ecosystem of biomolecular models has followed, including Boltz, Chai, OpenFold and Protenix. Yet commercial access remains uneven. AlphaFold 3, launched in 2024, broadened the field from protein folding to biomolecular interactions, while its public server remains off-limits for commercial use. DeepMind’s commercial engine sits inside Isomorphic Labs.

Drug developers are thus left with a choice between running open models on their own GPU clusters and pipelines, or renting access through infrastructure and workflow providers such as Tamarind Bio, Rowan or NVIDIA BioNeMo.

Boltz Bio is aiming at the opening with a new API that includes BoltzProt-1, a protein design pipeline, and BoltzMol-1, its first small-molecule hit-discovery pipeline. The company is also leaning into a newer interface for drug discovery: coding agents.

At its San Francisco launch event, co-founder and CEO Gabriele Corso said that, over the past few months, Boltz’s own scientists, chemists and protein engineers had been reaching the company’s models through coding agents. The new release takes that internal workflow public through an API, SDKs and integrations with Claude Code, Codex and Gemini CLI. In announcing the launch, the company put the bet more sharply: “We built this for agents as much as for people.”

The API is designed to be easy to use, Corso said. “The backend team led [the work] running thousands of [predictions], so anything you run will run very quickly,” he added. “We managed to make it at the cost of, or lower than, running [it yourself]. The IP stays yours.” Jeremy Wohlwend, Ph.D., co-founder and CTO at Boltz noted: “We spent quite a bit of time trying to figure out what the right interface to those agents would be.”

In a test for this story, we put that pitch to work. Running the Boltz API through Claude Code and its Boltz connector, we asked Claude in plain English to carry out a small hit-discovery screen using a handful of commercially available compounds against the EGFR kinase, a well-characterized cancer target.

The agent installed and authenticated the command-line tool on its own, refused to fabricate either the protein sequence or the compound library when asked, estimated the cost before spending anything and returned ranked structures in about three and a half minutes.

The test also revealed that with Boltz’s recommended filters enabled, the screen had discarded several of our most drug-like compounds before scoring any of them. The test used Boltz’s default “recommended” medicinal-chemistry filters plus a Lipinski and a PAINS filter.

Before submitting, the agent ran a cost estimate and reported the number. It estimated $0.025 per molecule, $0.20 for the eight-compound demo. Once I confirmed, the agent submitted the job, polled for completion in the background, downloaded the structures and scores, and ranked the results. Because Boltz only bills for compounds it actually scores, the run cost about $0.10 rather than the $0.20 quoted. From start to finish, it took about 3.5 minutes for the job.