Tamarind

Tamarind Bio

Tamarind Bio is a cloud platform that runs computational biology tools — structure prediction, protein and antibody design, docking, binding-affinity, MSA generation, and molecular dynamics — on managed GPUs. Users submit sequences or structures and get back predicted structures, designs, and biophysical scores, without provisioning their own hardware. It exposes hundreds of tools (AlphaFold, Boltz-2, Chai-1, RFdiffusion, ProteinMPNN, BoltzGen, ESMFold2, DiffDock, Autodock Vina, and many more) through one uniform job API.

Official docs: app.tamarind.bio/api-docs · platform UI at app.tamarind.bio

Canonical sources — fetch these, don't rely on a stale copy

Tamarind publishes live, machine-readable sources. Prefer fetching them at runtime over trusting any hardcoded list — tool names, schemas, and endpoints change frequently:

• https://app.tamarind.bio/llms.txt — LLM index: links to the spec, API docs, and MCP guide.
• https://app.tamarind.bio/openapi.yaml — OpenAPI 3.0 spec for the 8 core job endpoints (submit-job/-batch, jobs, result, upload, files, delete-job/-file; auth ApiKeyAuth). Fetch it for those exact shapes. Discovery/management endpoints (/tools, /usage-statistics, pipelines, …) aren't in it — use the MCP/REST discovery tools for those.
• https://docs.tamarind.bio/llms.txt — documentation index; every page has a .md form (e.g. docs.tamarind.bio/tamarind/batch.md, /tamarind/api.md, /tamarind/pipelines.md).
• Live tool discovery — GET /tools (REST) or MCP getAvailableTools + getJobSchema(jobType) are the source of truth for what tools exist and their parameters.

This skill teaches the surface + the non-obvious behaviors those sources don't spell out (see the reference files). When in doubt about a shape, fetch openapi.yaml.

When to use this skill

Use Tamarind when the user wants to:

• Predict structure of a protein, complex, or protein-ligand system (AlphaFold, Boltz-2, Chai-1, ESMFold2, Chai/Boltz cofolding)
• Design proteins or binders (RFdiffusion, BoltzGen, BindCraft, ProteinMPNN/LigandMPNN inverse folding)
• Design or characterize antibodies/nanobodies (sequence generation, humanization, developability, immunogenicity)
• Dock small molecules to a protein (DiffDock, Autodock Vina) or predict binding affinity
• Generate MSAs for downstream folding
• Run molecular dynamics or other biophysical workflows on managed GPUs
• Batch-screen many sequences or designs through the same tool
• Chain tools into pipelines (e.g. design → fold → score) using the output of one job as the input of the next

This skill is the right fit when the work should run on Tamarind's managed cloud rather than on a local install. For purely local cheminformatics or one-off sequence I/O, use a local library (RDKit, BioPython) instead.

Access and authentication

1. Sign in at app.tamarind.bio and create an API key from the account/API settings.
2. Authenticate every REST request with the x-api-key header.
3. Never hardcode the key. Read it from the TAMARIND_API_KEY environment variable or a .env file (use python-dotenv). Never commit keys to source control.

Pricing: Every user gets 10 free jobs. For larger usage, contact info@tamarind.bio to purchase a subscription.

export TAMARIND_API_KEY="your_api_key"
# List available tools
curl https://app.tamarind.bio/api/tools \
  -H "x-api-key: $TAMARIND_API_KEY"

Base URL: https://app.tamarind.bio/api/

There is no official Python SDK — the PyPI package named tamarind is an unrelated Neo4j tool. Do not pip install tamarind. Write plain requests calls against the REST API (the endpoint shapes are in openapi.yaml), or use the MCP server for agent hosts.

Two ways to call Tamarind

MCP server (best for AI agents)

Tamarind hosts an MCP server at https://mcp.tamarind.bio/mcp (API-key auth via the X-API-Key header). When your agent host supports MCP, prefer it — the tools mirror the REST API with agent-friendly schemas:

• listModalities() / listTags() — the live filter vocabulary (molecule type / function) with labels + tool counts; call these to learn valid modality/function values instead of hardcoding
• getAvailableTools(modality?, function?, search?, custom?) — discover tools (category/tag are deprecated aliases still honored)
• getJobSchema(jobType) — exact parameter schema for a tool, plus an exampleJob starting payload (validate it before submitting)
• validateJob(jobName, type, settings) — dry-run validation before submitting
• submitJob(jobName, type, settings) / submitBatch(batchName, type, settings[], jobNames[])
• getJobs(jobName?, batch?, limit?, includeSequences?) — list/inspect jobs and statuses (the bulky per-job input blob is omitted by default; pass includeSequences=true to keep it)
• getJobLogs(jobName) — fetch output logs for debugging
• listJobFiles(jobName) — list output files (returns s3Path for chaining)
• getResult(jobName, fileName?) — download results
• uploadFile(filename) — presigned upload URL; or uploadFileContent(filename, content, encoding?) to send file content through MCP when the host can't reach S3 (sandboxed agents)

Scope note: MCP query tools (getJobs, getResult, listJobFiles, …) are scoped to the authenticated account.

REST API (universal)

Use plain HTTP with requests — the endpoint shapes are in openapi.yaml. The core loop is below; references/workflows.md has full recipes.

Core workflow

Always follow discover → schema → validate → submit → poll → results. Do not hardcode tool names or settings — the catalog changes frequently.

import os, time, requests

BASE = "https://app.tamarind.bio/api"
HEADERS = {"x-api-key": os.environ["TAMARIND_API_KEY"]}

# 1. Discover tools. REST /tools returns the full list; filter client-side.
tools = requests.get(f"{BASE}/tools", headers=HEADERS).json()
alphafold = next(t for t in tools if t["name"] == "alphafold")

# 2. Get the exact schema for the chosen tool.
#    REST: each /tools entry already includes its inline `settings` schema
#          (parameter list) — find the entry whose name == your job type.
#    MCP:  getJobSchema(jobType) returns the same per-tool detail.

# 3. Submit a job. `settings` is tool-specific — match the schema exactly.
payload = {
    "jobName": "my-alphafold-run",          # ^[a-zA-Z0-9_-]+$, <=100 chars, unique
    "type": "alphafold",
    "settings": {
        "sequence": "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG",
        "numRecycles": 3,
    },
}
resp = requests.post(f"{BASE}/submit-job", headers=HEADERS, json=payload)
resp.raise_for_status()   # 200 ok; 400 bad request; 403 budget exceeded; 401 unauthorized

# 4. Poll for completion.
#    NOTE the response shape: GET /jobs?jobName=<name> returns the job ROW
#    directly (no "jobs" wrapper); the list query (no jobName) returns
#    {"jobs": [...]}. Don't index ["jobs"][0] on the by-name response.
while True:
    job = requests.get(f"{BASE}/jobs", headers=HEADERS,
                       params={"jobName": "my-alphafold-run"}).json()
    if job["JobStatus"] in ("Complete", "Stopped", "Deleted"):
        break
    time.sleep(30)

# 5. Retrieve results. POST /result returns a presigned URL *string*;
#    GET that URL to download the actual results zip (two-step).
url = requests.post(f"{BASE}/result", headers=HEADERS,
                    json={"jobName": "my-alphafold-run"}).text.strip('"')
open("my-alphafold-run.zip", "wb").write(requests.get(url).content)

For the agentic version of this loop using MCP tools, and for richer examples, see references/workflows.md.

Discovering tools

The catalog has hundreds of tools. Always enumerate at runtime — never rely on a hardcoded list.

REST GET /tools returns the full list (it does not filter server-side); each item is {name, displayName, github, paper, description, settings} where settings is that tool's inline parameter schema. Filter client-side:

tools = requests.get(f"{BASE}/tools", headers=HEADERS).json()   # a list
boltz = [t for t in tools if "boltz" in t["name"].lower()]

Note: both surfaces return one row per tool name — REST /tools and MCP getAvailableTools are both deduplicated (the MCP keeps the newest tool version), so a name match returns a single row.

MCP getAvailableTools(search=..., modality=..., function=...) filters server-side and adds categories/tags per tool (category/tag are deprecated aliases of modality/function, still honored). Don't hardcode the vocabulary — it drifts. Get the live values from listModalities() / listTags() (each returns value, label, description, and toolCount), or read the availableCategories / availableTags facet arrays returned on every getAvailableTools response. Modalities are molecule types (protein, antibody, peptide, small-molecule, nucleic-acid, …); functions are what a tool does (structure-prediction, binder-design, protein-ligand-docking, …).

A representative set of widely-used tools (verify with /tools): alphafold, boltz (Boltz-2), chai (Chai-1), esmfold / esmfold2, rfdiffusion, proteinmpnn, ligandmpnn, boltzgen, bindcraft, diffdock. See references/tool_catalog.md for the full category/tag map and how to read tool metadata.

Choosing the right tool

The catalog has many tools per task; don't hardcode a favorite — filter by function (and modality), then read each candidate's description and match it to the user's actual goal (input you have, output you need, constraints like speed or "no MSA"). The description and tags fields are the public "what it's for" signal; let them, plus validateJob, drive the pick. Quick orientation by task:

• Fold a single protein / complex (function=structure-prediction): the AlphaFold3-class reproductions — boltz/chai/openfold/protenix/intfold — are the accurate default for everything, including protein-only systems; they also handle nucleic-acid + small-molecule complexes, so reach for them whenever a ligand/RNA/DNA is part of the system (and boltz adds binding-affinity). alphafold (AF2) remains a solid choice for monomers + multimers (join chains with :). esmfold is single-sequence (no MSA) and fast — reach for it when you want speed and have no MSA; esmfold2 is newer and conditions on an MSA by default (its model setting offers a faster single-sequence mode). Specialized folders exist for antibodies (abodybuilder, immunebuilder), cyclic peptides (highfold), and conformational ensembles (afcluster, alphaflow) — filter and read descriptions.
• Design a binder (function=binder-design): bindcraft (de novo miniprotein binders) and boltzgen (binders for protein and small-molecule targets, incl. nanobodies/antibodies/peptides) are the go-to de novo binder tools; rfdiffusion also does binder design and is the pick for motif scaffolding / diversifying an existing backbone. Antibody-specific generators live under function=antibody-design.
• Design sequence for a known backbone (function=inverse-folding): proteinmpnn (general), ligandmpnn (ligand-aware), plus thermostable/soluble/antibody MPNN variants. Inverse folding takes a structure and emits sequences — fold them back to verify (see chaining).
• Dock a small molecule (function=protein-ligand-docking): prefer boltz/chai — they co-fold the ligand into the complex and predict the bound structure rather than docking into a fixed receptor; reach for autodock-vina when you need fast, large-scale screening against a known pocket.
• Predict binding affinity (function=binding-affinity) or generate an MSA (search msa) — filter and read.

When the user names a specific tool, evaluate that one and sanity-check the alternatives in its tag group — a faster or more appropriate sibling often exists. When unsure, getJobSchema/validateJob to confirm a candidate actually accepts the input you have before committing.

Job settings, schemas, and validation

Each tool has its own settings schema. Fetch it before submitting:

• REST /tools entry: each settings param is a trimmed dict. Only name and required are always present; type, default, description, options appear only when relevant (≈60% have type) — so use param.get("type"), not param["type"]. The advanced gating keys (exclude, conditionals) are NOT in the REST response at all.
• MCP getJobSchema(jobType): the full schema, including exclude, conditionals, and bounds. Use MCP when you need to reason about those gating keys. (restrictOrgs is stripped on both surfaces — an org-gated param you can't use is simply omitted; see references/api_reference.md.)

Always validateJob (MCP) before submitting — it's the reliable guard. It runs the same validation as /submit-job without submitting, and surfaces the first missing/invalid field. Don't try to hand-derive which fields to strip from the schema keys (over REST you can't see them anyway) — let validateJob tell you. (The response may include a source field, e.g. "static-fallback" — an internal note on which schema source validated; valid: true/false is the signal you act on.)

validateJob echoes a normalized view of your settings with defaults filled in. Submit the same clean settings you validated; treat normalized as informational (it can carry defaults you didn't set, and for some tools platform-managed fields), so build your submit from your own settings rather than the normalized blob.

Sequences: amino-acid string; separate chains of a multimer with a colon (:), e.g. "MVLS...:EVQL...". Note that some tools (e.g. boltz, chai) require more than sequence — boltz also requires inputFormat (and accepts yamlFile/molecules). Always getJobSchema/validateJob to learn a tool's required fields; don't assume sequence alone suffices.

Platform-internal fields — never set these yourself; the platform owns them: submit_method, monomer_msa, msa. See references/api_reference.md for the full field-handling rules.

Surface consequential choices before submitting, don't default silently. When the request fully specifies what to run, proceed. But when it's open-ended, or when a setting materially changes the results, runtime, or cost (model/variant, number of samples or seeds, MSA on/off, GPU tier, batch size), present the meaningful options plus the default you'd otherwise apply and let the user pick before you submit — rather than choosing silently and reporting it after the job is queued. getJobSchema and validateJob's normalized show exactly which knobs you're filling in on the user's behalf, so you can flag the few worth a quick confirm. This matters most for batches, where one shared-settings choice multiplies across every job.

File inputs (PDB, CIF, SDF, …)

Tools with file parameters accept input three ways:

1. Upload first, then reference by bare filename. PUT /upload/{filename}, or MCP uploadFile → presigned URL → curl -X PUT -T file "<url>". If your host can't reach S3 (a sandboxed agent with no outbound network), use MCP uploadFileContent(filename, content, encoding?) to send the file's content through the MCP channel instead — text by default, encoding="base64" for binary. The object lands at the S3 key {email}/{filename}, but you reference it in settings by the bare filename only (e.g. "targetFile": "GLP1R_ECD.pdb") — the platform scopes it to your account automatically. Do NOT prefix the email: passing {email}/{filename} double-prefixes the lookup and submit-job 400s with "The following files have not been uploaded: <email>/<file>". Confirm the exact name the store registered with MCP getFiles(search=...) / REST GET /files (a flat list of bare names).
2. Reference a prior job's output by its path: JobName/path/to/file.ext (this is how you chain jobs — see below).
3. Inline content. Send the file's text content directly as the field value.

Foot-gun: for a file-typed parameter, a plain string value is treated as inline file content, not as a path to an existing object. To point at an already-uploaded file, use the bare filename (not the {email}/... S3 key) or, for a prior job's output, the JobName/... path form — not a bare string you expect to resolve to new content.

validateJob notes. The response may carry a source field (e.g. "static-fallback") — it labels how the tool's schema was resolved (built-in tools always report static-fallback), not whether the validator was reachable, so act on valid, not source. For file params: reference an uploaded file by its bare filename (above) — a bare name resolves to your account-scoped object, whereas an email-prefixed string can be read as inline content and fail the file-type check ("... must contain ATOM records"). And passing inline file content makes validateJob upload it synchronously before validating, which can be slow; prefer referencing an uploaded file by name (above). If a dry-run is slow, skip it and let submit-job validate.

Chaining jobs into pipelines

A finished job's output becomes the next job's input — no download/re-upload. Match the input type the next tool actually wants: a sequence-design tool (ProteinMPNN) emits sequences, so you fold them by passing each as a sequence; a tool that takes a file parameter takes a path.

The cleanest design→fold chain is the MCP submitBatch(fromJob=...), which reads a completed design job's generated sequences and folds each as one job:

# ProteinMPNN designs sequences -> fold every one with AlphaFold, one call:
submitBatch(batchName="verify-designs", type="alphafold", fromJob="my-proteinmpnn-job")

For a file input (e.g. a tool that takes a .pdb/.cif), reference a prior job's output by the path form JobName/path/to/file.ext in that file parameter. Two cautions, both confirmed by validation: (1) match the parameter's required file type — e.g. AlphaFold's templateFiles accepts only .cif and is a list, and is gated behind templateMode: "custom"; (2) templateFiles is for structural templates, not for "fold this designed sequence" — to fold a sequence, pass sequence. Always getJobSchema/validateJob to confirm a file param's type/conditions before chaining into it.

To discover a job's exact output paths, use MCP listJobFiles(job1) — it returns each file's s3Path, usable directly in the next submitJob. (The REST GET /files lists your account's uploaded files as a flat name list; it does not enumerate a job's outputs.) Tamarind also supports saved pipelines: build one in the UI, then drive it with /run-pipeline ({pipelineName, initialInputs, inputs}) or define stages[] inline via /submit-pipeline (each stage names a task + toolSettings, using "pdbFile": "pipe" to thread one stage's output into the next). See references/workflows.md.

Batch submission

Submit many jobs of the same tool in one call. The Python form uses parallel settings[] and jobNames[] arrays (same length, up to 100):

requests.post(f"{BASE}/submit-batch", headers=HEADERS, json={
    "batchName": "egfr-binder-screen",
    "type": "alphafold",
    "jobNames": ["seq1", "seq2", "seq3"],
    "settings": [{"sequence": "..."}, {"sequence": "..."}, {"sequence": "..."}],
    # optional: "maxRuntimeSeconds": 3600, "weightedHoursBudget": 100,
    # (some accounts also accept an optional "gpuType" — confirm with support)
})

Poll the batch parent on batchStatus, not subjob JobStatus. A batch creates a parent job (Type: "batch") plus subjobs. Subjobs flip to Complete as soon as they finish computing, but the batch then spends a few minutes aggregating results into the final downloadable output. Fetch the parent by name and watch batchStatus:

import time
while True:
    # ?jobName= returns the parent ROW directly (no "jobs" wrapper)
    parent = requests.get(f"{BASE}/jobs", headers=HEADERS,
                          params={"jobName": "egfr-binder-screen"}).json()
    bs = parent.get("batchStatus")
    if bs == "Complete":
        break
    if bs in ("Stopped", "AggregationFailed"):
        raise RuntimeError(parent.get("AggregationError", bs))
    time.sleep(15)   # Running / Aggregating -> keep waiting
# When Complete, the parent carries a presigned `resultUrl` and a `statuses`
# subjob tally ({Complete, Running, In Queue, Stopped}).
open("batch.zip", "wb").write(requests.get(parent["resultUrl"]).content)

Add includeSubjobs=true to GET /jobs?batch=<name> to list per-subjob rows.

Job status lifecycle

Single jobs report JobStatus; batch parents report batchStatus (poll that for batches — see above).

Completed jobs carry a Score (tool-specific metrics, e.g. pLDDT/pTM/ipTM for folding) and WeightedHours. Treat Complete/Stopped/Deleted (and AggregationFailed for batches) as terminal; poll on a 15-30s interval. Break your poll loop on any terminal status, not just Complete/Stopped — a job that goes Deleted mid-poll would otherwise loop forever. For a Stopped job, fetch getJobLogs(jobName) to see why. WeightedHours is the usage unit billed per job; cap a batch with weightedHoursBudget, and a 403 on submit means a budget was hit (see references/api_reference.md and the /usage-statistics endpoint).

Error handling

Reference files

The openapi.yaml spec is the source of truth for endpoint shapes; these files add the behaviors and gotchas the spec doesn't spell out:

• references/examples.md — validated settings payloads per common tool (alphafold/boltz/diffdock/autodock-vina/proteinmpnn/batch), a copy-paste self-check, the "what fails and the exact error" list, and output-shape notes. Start here for a working payload.
• references/api_reference.md — endpoint quick-reference + the non-obvious shapes: /jobs by-name returns a bare row (not {jobs:[...]}), /result is a two-step download, batch parents poll on batchStatus, /files is a flat name list, the settings field-handling rules.
• references/tool_catalog.md — category/tag map, how to read tool + parameter metadata, common tool families.
• references/workflows.md — end-to-end recipes: fold a sequence, validate-before-submit, upload + reference a file, design→fold chaining, batch screen with aggregation polling, usage stats, pagination, and the non-blocking submit-now/check-later pattern for long jobs.

Source: K-Dense-AI/scientific-agent-skills → skills/tamarind/SKILL.md