k-dense-ai/scientific-agent-skills/adaptyv

Adaptyv Bio Foundry API

Use this skill to turn protein sequences into experimentally measured data through Adaptyv's cloud laboratory. Confirm the target construct, assay conditions, controls, and replicate plan before submitting a batch; turnaround depends on the experiment.

Reviewed against the deployed OpenAPI schema (info.version: 0.0.2) and official SDK 0.1.0 at commit cdf207819ed5a58e0c127453626d2bce125c8064. The schema version alone does not identify all changes. See the endpoint reference for current request/response details and discrepancies in upstream examples.

Connect

Base URL: https://devs.adaptyvbio.com/api/v1. The schema is at /openapi.json; never append that filename to normal endpoint requests.

Create a bearer token in Foundry under Organization → Settings → Tokens. Use Viewer for reads and Member for experiment writes. Read credentials from ADAPTYV_API_KEY; the documentation's FOUNDRY_API_TOKEN is an alternative variable name for the same token. If using a project .env, explicitly call python-dotenv.load_dotenv() before SDK setup; the SDK does not load that file automatically. Never print or commit tokens.

# ADAPTYV_API_KEY is already set in the environment.
curl --fail-with-body 'https://devs.adaptyvbio.com/api/v1/targets?limit=3' \
  -H "Authorization: Bearer $ADAPTYV_API_KEY"

Resource endpoints require bearer authentication. The schema and liveness endpoint GET /info/health are public. GET /whoami reports the active organization and permissions; check it when account scope is unclear.

Python SDK

Install the reviewed revision in your project environment (requires Python 3.11+):

uv pip install "git+https://github.com/adaptyvbio/adaptyv-sdk.git@cdf207819ed5a58e0c127453626d2bce125c8064"

Import from adaptyv, not adaptyv_sdk. FoundryClient requires explicit api_key and base_url; the lab convenience object reads ADAPTYV_API_KEY and ADAPTYV_API_URL. Set the latter to the base URL above to override the SDK's older hostname default.

The following examples are illustrative for authenticated use. Their request construction and parsing were checked with mocked HTTP responses, not a paid lab run. Replace environment inputs with real, reviewed sequence data and catalog IDs.

Browse, estimate, create a draft

import json
import os
from pathlib import Path
from adaptyv import FoundryClient
client = FoundryClient(
    api_key=os.environ["ADAPTYV_API_KEY"],
    base_url="https://devs.adaptyvbio.com/api/v1",
)
# Review vendor, catalog number, construct, and pricing before selecting an ID.
targets = client.targets.list(search="EGFR", selfservice_only=True, detailed=True)
for target in targets.items:
    print(target.id, target.name, target.vendor_name, target.catalog_number)
# candidates.json is a map of unique names to full amino acid sequences.
sequences = json.loads(Path("candidates.json").read_text())
spec = {
    "experiment_type": "screening",
    "method": "bli",
    "target_id": os.environ["ADAPTYV_TARGET_ID"],
    "sequences": sequences,
    "n_replicates": 3,
}
# SDK takes the spec itself; REST takes {"experiment_spec": spec}.
estimate = client.experiments.cost_estimate(spec)
if estimate.breakdown is None:
    raise ValueError("Incomplete estimate: review warnings and obtain a full quote")
print("Estimated USD cents, excluding VAT:", estimate.breakdown.total_cents)
# After the batch and estimated cost have been reviewed:
exp = client.experiments.create(name="EGFR binder screen batch 1", experiment_spec=spec)
experiment_id = exp.experiment_id
client.close()

create(name=..., experiment_spec=...) does not accept a single REST-body dictionary. cost_estimate(spec) wraps the spec itself; do not wrap it a second time. For an affinity experiment, the SDK additionally requires explicit antigen_concentrations, even though REST supplies a default when omitted.

Submit, then inspect and confirm the quote

With a configured client and the saved experiment_id:

client.experiments.submit(experiment_id)
# Quote generation is asynchronous. Retry get_quote with a bounded timeout if it
# returns 404 after successful submission; a draft has no forthcoming quote.
quote = client.experiments.get_quote(experiment_id)
print(quote.amount_total, quote.currency, quote.expires_at)
# Once the quote is accepted within the user's authorized scope:
accepted = client.experiments.confirm_quote(experiment_id)
print(accepted.invoice_id, accepted.hosted_invoice_url)

Confirmation creates/finalizes an invoice; it does not settle payment. Use the returned hosted invoice URL or the current REST payment pointer. SDK 0.1.0 does not expose invoice payment, organization webhooks, or whoami helpers.

Avoid the reviewed SDK's lab.experiment(target="EGFR") shortcut: it passes only UUID-shaped targets and does not resolve a target name. auto_confirm=True and lab.confirm_experiment() call submission, not the quote-confirm endpoint. Use the explicit client methods above. The decorator also supplies method for non-binding experiments, which the current API rejects.

Retrieve every result without losing fields

The SDK's generated models predate some live result fields and may discard them. Use REST JSON when archiving results, target references, or newer kinetic fits:

import json
import os
from pathlib import Path
import httpx
experiment_id = os.environ["ADAPTYV_EXPERIMENT_ID"]
with httpx.Client(
    base_url="https://devs.adaptyvbio.com/api/v1",
    headers={"Authorization": f"Bearer {os.environ['ADAPTYV_API_KEY']}"},
    timeout=30,
) as api:
    response = api.get(f"/experiments/{experiment_id}")
    response.raise_for_status()
    experiment = response.json()
    if experiment["results_status"] != "all":
        raise RuntimeError("Results are not complete; inspect status before analysis")
    results, offset = [], 0
    while True:
        response = api.get(f"/experiments/{experiment_id}/results",
                           params={"limit": 100, "offset": offset})
        response.raise_for_status()
        page = response.json()
        results.extend(page["items"])
        offset += len(page["items"])
        if not page["items"] or offset >= page["total"]:
            break
Path("foundry-results.json").write_text(json.dumps(
    {"experiment": experiment, "results": results}, indent=2
))

Archive raw data packages when available, plus sequence/target identities, assay method, units, conditions, replicate measurements and fit quality. Null kinetic values are missing/unresolved measurements, not zero. A screening binding call is not a measured affinity. Compare KD values only under compatible assay conditions; the schema's kd_mean averages strong-binding replicates and is not an unbiased summary of every tested replicate.

Experiment types and validation

Inapplicable fields are rejected. antigen_concentrations is affinity-only, in nM; REST's default is [1000.0, 316.2, 100.0, 31.6, 0.0]. parameters holds optional assay settings; coordinate any nonstandard configuration with the laboratory.

Sequences accept full strings or rich entries such as {"candidate": {"aa_string": "EVQLVESGGGLVQPGGSLRLSCAAS", "control": false}}. Use the 20 standard amino acid letters; inputs are case-insensitive and stored uppercase. Colons separate chains. Ellipses are not valid sequence characters. Rich creation metadata is constrained by SequenceMetadata, not arbitrary JSON: use SingleChain, ScFv, FAB, or IgG; ScFv needs VH and VL, FAB needs framework_regions.ch and .cl. The SDK accepts the schema's Portal-style enum values. Add sequences through POST /sequences only while status is draft.

Lifecycle and updates

Wire status values are lowercase snake case:

draft -> waiting_for_confirmation -> quote_sent -> waiting_for_materials
      -> in_queue -> in_production -> data_analysis -> in_review -> done

canceled is also possible. results_status is independently none, partial, or all. Inspect it before treating a result set as complete. Most experiment PATCH fields are editable in draft or in_review; webhook_url remains editable at any status. A sequence PATCH replaces the sequence list; POST /sequences appends to a draft.

Create-time REST flags skip_draft and auto_accept_quote enable automation; auto_accept_quote implies skip_draft and requires full pricing. They can commit the batch to laboratory processing and billing, so use them only within the user's approved experiment/budget scope. The SDK's public create() method does not expose auto_accept_quote, webhook_secret, or the payment selector; use the REST contract in the reference when these fields are needed.

For signed delivery, set a per-experiment webhook_url and webhook_secret (minimum 32 characters) at creation, or register an organization webhook. Store the secret when setting it; it is write-only. An experiment webhook overrides organization delivery. Verify X-Adaptyv-Signature (sha256=<hex>) as HMAC-SHA256 over the raw body with constant-time comparison. Events are experiment_update customer-facing updates, not a complete stream of status transitions. Deduplicate by data.update_id; delivery_id identifies an attempt. Return 2xx; network failures and 5xx retry up to three times, while 4xx is permanent. Continue to poll status/results if delivery is missing.

Lists, filters and errors

Paginated endpoints return {items, total, count, offset}; use limit 1–100 (default 50) and offset (default 0). Fetch all pages. Query support varies by endpoint: sequence lists lack filter, token lists only paginate, and organization webhooks return an unpaginated array. See the endpoint reference.

Where supported, filters use eq(field,value), neq, gt, gte, lt, lte, contains, between, in, is_null, is_not_null, combined with and, or, not. Example: and(gte(created_at,2026-01-01),eq(status,done)). Sorting accepts asc(field) / desc(field) (up to eight comma-separated terms), -field / +field, or field:asc / field:desc. Use the HTTP client's query encoder. Advanced expressions support at(field,key) and casts (float, int, text, timestamp, date), subject to each endpoint's allowed fields.

Structured errors generally include error and request_id; record the x-request-id header for support. Handle non-JSON transport failures too. Check status before retrying mutations after an ambiguous timeout: the server may already have created the experiment or accepted a quote. The SDK retries 429/5xx responses; its retries do not establish mutation idempotency.

Token attenuation only narrows permissions. POST /tokens/revoke revokes the calling token's root family, including siblings descended from that root, even when authenticating with an attenuated token.

Citing Scientific Agent Skills