🧬boltz-protein-design

Workflow

If boltz-api is missing from PATH, use boltz-cli-setup for install/update guidance before retrying. If a command reports missing or expired authentication, use boltz-cli-setup to start boltz-api auth login --device-code before retrying; do not ask permission first. If the agent host sandbox blocks boltz-api install/auth/API calls, use boltz-cli-setup to request the host sandbox bypass/escalation needed for user-wide CLI install, browser login, credential storage, temp files, or API access before retrying.

Use this skill when the user wants de novo protein / peptide / antibody / nanobody binders.

1. Decide on target exploration first (new targets). For a new target where the user hasn't already fixed the binding site and crop, your first action — before authoring a payload, normalizing the target, or running estimate-cost — is to raise the choice between a target-exploration pass and designing directly, with a recommendation for this target:

   • Unknown site, or a multi-domain / large target → recommend exploration (it scouts different input configurations for generation, ≈50 designs each, and finds the best before a full run).
   • A well-characterized site → it's fine to recommend going (mostly) direct, perhaps with a quick check of whether conditioning on the epitope beats letting the model find its own spot. State this plainly, as part of a conversation with the user about their target and goals, and let them choose.

   Phrase it as a question that works with the user (they may know their target's biology), e.g.:

   "This is a fresh target — I'd suggest a quick exploration pass that scouts a few framings and picks the best before a full run. Or, if you already know the site and crop, we can design directly. Which would you like?"

   Do not mention a campaign size or tier here — not even folded into this opening approach question. The full-run size is settled later, after the scouting runs pick a winner (its yield informs the tier), so don't ask it up front when exploration is on the table. If the user opts into exploration — or has already said they want to explore / let the design find its own epitope — read references/target-exploration.md, follow it, then resume at step 8 with the chosen framing and recommended num_proteins. If they want to design directly, continue below.

2. Normalize the target (same shape as protein-screen): structure_template if a CIF/PDB is available, else no_template.

3. Pick the binder_specification variant. Supported variants include:

   • boltz_curated — recommended default for antibody and nanobody design. Boltz selects from maintained scaffold/template lists (binder: boltz_antibody or boltz_nanobody).
   • structure_template — redesign motifs in an existing binder scaffold (CIF + design_motifs with replacement / insertion segments).
   • no_template — generate from the sequence DSL (fixed residues + designed segments like 5..10 or 8).

4. For antibody or nanobody requests, ask before authoring the payload: "I recommend Boltz's curated antibody/nanobody scaffolds for this. Do you want the curated default, or do you have custom scaffold structures/CDR motifs to use?" If the user picks curated, use type: boltz_curated; if they want custom scaffold control, use type: structure_template.

5. Pick modality: peptide, antibody, nanobody, or custom_protein for structure_template and no_template (use custom_protein for a "miniprotein" or generic "protein binder"). If the user already named the modality, take it as given — don't ask again. Do not include modality on boltz_curated; use binder instead.

6. Pick num_proteins — see Run sizing. Valid range is 10 to 1,000,000 (server rejects outside it); 10 is the hard floor but it is a test size, not a campaign. When the user has not given a count, propose a campaign tier (default 50,000), not the floor.

7. Supported optional features include rules such as excluded amino acids, excluded sequence motifs with X wildcards, and max hydrophobic fraction. Add rules only on request; read references/api.md for exact shapes and examples.

8. Author the payload YAML or JSON, then run estimate-cost and apply the spending gate (Always Do This) before start. (Cost model — tiered by total complex length, estimate-cost is the only source: see ## Cost in api.md.)

9. start to submit. Capture the ID.

10. Launch download-results with the agent runtime's background/non-blocking command facility. In Claude Code, use Bash with run_in_background: true. In Codex, run download-results as a foreground shell command with yield_time_ms: 1000; if Codex returns a session_id, keep it for optional same-thread polling, but treat download-status plus the run directory as the durable source of truth. In Codex app/desktop runtimes that expose same-thread heartbeat automations, create a heartbeat that checks download-status periodically and posts a concise completion or failure update when the download reaches a terminal state. After launching the downloader, always report the job ID, run name, and output directory. Include the next check cadence if the heartbeat was created; otherwise include the download-status command.

11. Rank from <output-root>/<run-name>/results/index.jsonl by binding_confidence descending. Use iptm and min_interaction_pae as tiebreakers. optimization_score is not emitted for this endpoint. Read references/results.md for output layout and metric details.

Run sizing

De novo design is a generate-and-filter campaign: you make many binders and keep the rare good ones, so a real run is large. Do not anchor on the num_proteins floor of 10 — that is only useful for a quick setup test. When the user names a count, honor it (≥10). When they do not, explain the tiers and propose one:

Tier	num_proteins	When
Small screen	20,000	Quick look / tight budget
Medium (recommended)	50,000	Default for a real campaign
Large	100,000	Hard target / maximal coverage

Present the tiers as design counts, not dollars: don't put a price next to a tier unless estimate-cost returned it — run estimate-cost on the tier the user leans toward and show that figure, and never extrapolate one estimate across the others. Then apply the spending gate (Always Do This) before submitting. If the setup is unproven, suggest a small test run (tens of designs) or the full target-exploration pass first.

Command Pattern

# Replace placeholders with concrete absolute paths before running.
# Use a short descriptive run name, for example: protein-design-<modality>-<target>-v1

boltz-api protein:design estimate-cost \
  --input @yaml:///absolute/path/payload.yaml

boltz-api protein:design start \
       --idempotency-key "<run-name>" \
       --input @yaml:///absolute/path/payload.yaml \
       --raw-output --transform id

# Copy the printed job ID into this command, then launch it in the agent
# runtime's background/non-blocking mode.
# Claude Code: Bash with run_in_background=true.
# Codex: foreground shell command with yield_time_ms=1000; keep the returned session_id if one is provided.
# Do not append "&" or use nohup in Codex.
boltz-api download-results \
  --id "<job-id-from-start>" --name "<run-name>" \
  --root-dir "/absolute/path/boltz-experiments" \
  --poll-interval-seconds 60

Payload keys are num_proteins, target, binder_specification — API body field names.

Always Do This

• For a new target, your first move is the step-1 conversation — never jump straight to a payload. Establish what the binder is for and whether the user has already fixed the binding site/crop, then recommend a target-exploration pass vs. designing directly and let them choose (step 1). Only normalize the target and author a payload after that. If they've already fixed the site/crop or explicitly want to design directly, proceed.
• Enforce 10 <= num_proteins <= 1,000,000 before calling estimate-cost (server rejects outside that range), but 10 is the floor, not a campaign — see Run sizing and propose a tier (default 50,000) when the user gives no count.
• Spending gate — explicit go-ahead before every start. start spends real money. A plan you already described, an earlier phase's approval, or a cost that looks "trivial" are not authorization — even a cheap run needs a fresh yes. Run estimate-cost, show the estimated_cost_usd it returns (summed for a batch), and wait for the user to say go. This holds even when tool calls are pre-approved (accept-edits / auto-accept / bypass modes) — there you are the only cost gate. Never quote or assume a dollar figure you didn't get from estimate-cost (cost model: ## Cost in api.md).
• For antibody or nanobody design, recommend binder_specification.type: boltz_curated and ask the user to confirm they do not want custom scaffold/CDR control before building the payload. Use binder: boltz_antibody for antibody/Fab requests and binder: boltz_nanobody for nanobody/VHH requests.
• Residue indices are 0-based everywhere (design_motifs.start_index/end_index, after_residue_index, epitope_residues, flexible_residues, bonds, constraints).
• For CIF/PDB bytes, use @data:///abs/path/file.cif inside structure.data. Don't use bare @path.
• Sequence DSL for designed_protein.value: uppercase letters = fixed residues; integer N = exactly N designed residues; MIN..MAX = variable-length designed segment. Examples: "20", "5..10", "ACDE8GHI", "MKTAYI5..10VKSHFSRQ".
• Keep payload field names exactly as the API body names shown in references/api.md.
• Use absolute paths for the output root, payload files, and embedded target files. Do not cd into the run directory for follow-up commands; pass the same --root-dir and use absolute paths so later relative paths do not drift.
• Prefer one merged top-level payload via --input @yaml:///absolute/path/payload.yaml or @json:///absolute/path/payload.json for estimate-cost and start. Keep --idempotency-key and --workspace-id top-level; if they also appear inside --input, the top-level flags win.
• Direct object flags still work as overrides, such as --target @yaml:///absolute/path/target.yaml or --binder-specification @json:///absolute/path/binder.json. Piped YAML / JSON on stdin also works, but it must use API body field names. Use the same slug for both --idempotency-key and --name.
• In permission-gated agents such as Claude Code, keep each Boltz call as a top-level command that starts with boltz-api. Prefer concrete arguments over sh -c, inline environment assignments, aliases, wrapper scripts, loops, or pipelines around the boltz-api invocation unless the user already allowed that exact command form. Use --raw-output --transform id, read the printed ID, then paste that literal ID into the next download-results command.
• Prefer the agent runtime's background/non-blocking command mode for download-results. In Codex specifically, keep download-results in the foreground and set the shell tool yield to 1000 ms; Codex will return a session_id if the command is still running. Do not append & or use nohup in Codex because the tool runner may clean up shell-backgrounded descendants before .boltz-run.json is fully written.
• After the background/session starts, do not manually wait on it or run ad hoc polling loops. Wall-clock time scales roughly with num_proteins: under 100 often finishes in a few minutes, 100-1,000 may take several minutes to tens of minutes, and larger runs can take longer or hours depending on inputs and system load. Don't quote a fixed duration. --poll-interval-seconds 60 is a sensible downloader default. download-results emits JSONL progress on stderr by default; add --progress-format text --verbose only when you explicitly want human-readable logs.
• In Codex app/desktop runtimes with same-thread heartbeat automation support, schedule a heartbeat after launching download-results. The heartbeat should run boltz-api --format json download-status --name "<run-name>" --root-dir "/absolute/path/boltz-experiments" and stop once terminal. Choose cadence by num_proteins: under 100 -> every 1-2 minutes; 100-1,000 -> every 5 minutes; over 1,000 -> every 15 minutes. Post only material status changes or terminal completion/failure. Poll the saved session_id with an empty write_stdin only for interactive, user-requested progress checks. Never run a manual poll loop in the current turn.
• If the current host has no heartbeat automation support, do not claim an automatic next check. Report the job ID, run name, output directory, and the command needed to check download-status.
• If detached download needs to be restarted, re-run boltz-api download-results with the same --name "<run-name>" and the same --root-dir.
• Only add rules on explicit user request.

Escape Hatch

• Payload reference: https://api.boltz.bio/docs/api/resources/protein/subresources/design/methods/start/
• CLI flag names: boltz-api protein:design start --help

Read references/api.md for all binder_specification variants, motif shapes, sequence DSL, rules, modalities, and target variants. Read references/results.md after download when ranking designed binders or explaining outputs.

Outputs

Rank from results/index.jsonl after download-results; use references/results.md for local file layout, metric meanings, and the designed-binder entity type gotcha.