Tamarind CLI

Command-line interface for the Tamarind Bio platform. Discover tools, submit and monitor protein / nucleic-acid / small-molecule jobs, and download results — from your terminal, a script, CI, or an AI coding agent (Claude Code, Codex, …).

The CLI is a thin client over the same API the Tamarind MCP server uses, so the two stay in lockstep. See docs/architecture.md for how drift is prevented.

Install

uv tool install tamarind-cli      # or: pipx install tamarind-cli

Or with the bootstrap installer (uses uv, pipx, or Python/pip already on your PATH):

curl -fsSL https://app.tamarind.bio/cli/install.sh | sh

To track the repo directly instead of the PyPI release:

uv tool install "git+https://github.com/Tamarind-Bio/tamarind-cli"
# or: pipx install "git+https://github.com/Tamarind-Bio/tamarind-cli"

Releasing: tag a GitHub Release and the publish.yml workflow builds and uploads to PyPI via Trusted Publishing (configure the trusted publisher for tamarind-cli on PyPI first).

Authenticate

Get an API key from the Tamarind web app (Settings → API), then either:

export TAMARIND_API_KEY="sk_..."      # best for agents / CI
# or
tamarind auth login                    # stores it in ~/.tamarind/config.json
tamarind auth status

Quickstart

# 1. Find a tool
tamarind tools --function structure-prediction --modality protein
tamarind tools --search boltz

# 2. Inspect its parameters and grab a runnable example
tamarind schema boltz
tamarind schema boltz --example > job.yaml   # most tools ship an example; some don't

# 3. Validate, then submit
tamarind validate boltz --input job.yaml
tamarind submit boltz --input job.yaml --name my-run \
  --wait --timeout 3600 --download ./out

# 4. Monitor / fetch
tamarind jobs
tamarind status my-run
tamarind results my-run --download ./out

Set individual fields inline instead of a file:

tamarind submit boltz \
  --set inputFormat=sequence \
  --set sequence=MKTVRQERLKSIVRIL... \
  --name quick-fold

Python SDK

The package exposes typed Pipelines and Custom Tools resources. Pipeline runs use the same node-run terminology as the web app and public REST API:

from tamarind import Tamarind

with Tamarind() as client:
    run = client.pipelines.get_run("run-id")
    for node_run in run.node_runs:
        page = node_run.molecules(limit=100)
        print(node_run.label, page.items)

The Custom Tools resource checks archive-local safety, packages a folder, uploads it directly to object storage, and submits the digest-checked build request. The server remains authoritative for source readiness, Versions, logs, and the evolving config.json contract.

from tamarind import Tamarind
from tamarind.errors import CustomToolNotFoundError

with Tamarind() as client:
    try:
        tool = client.custom_tools.get("my-esmfold")
    except CustomToolNotFoundError:
        tool = client.custom_tools.create("my-esmfold", display_name="My ESMFold")

    result = tool.build("./my-esmfold", idempotency_key="release-2026-08-26")
    print(result.action)  # build, reuse_image, or unchanged
    version = result.version
    print(version.version)  # numbered version, e.g. v3
    if not version.terminal:
        version = version.monitor(timeout=1800, on_event=print)

build() returns a typed result describing what the request did and the durable Version it produced. It is convenience orchestration, not a durable request object. Reuse idempotency_key when retrying an ambiguous build response; the server returns the already admitted Version. If no key was supplied, fetch the tool's versions before retrying. Interrupting monitor() stops local monitoring; it does not cancel the remote build.

Exact Version operations use version.version, such as v3. If a mutation reports 412 Precondition Failed, refetch the affected Tool or Version, confirm the mutation is still desired, and retry using the refreshed resource.

The same lifecycle is available through the unified CLI. A typical release is:

tamarind custom-tools validate ./my-esmfold
tamarind custom-tools create my-esmfold --display-name "My ESMFold"
tamarind custom-tools build my-esmfold ./my-esmfold \
  --idempotency-key release-2026-08-31 --wait --timeout 1800
tamarind custom-tools versions my-esmfold
tamarind custom-tools publish my-esmfold v3

Build responses contain a numbered version, such as v3. Pass it to version, logs, cancel, and publish. A local wait timeout does not cancel the remote build; reattach with tamarind custom-tools version NAME VERSION --wait.

Test a completed Custom Tool version

After building, submit a test without publishing the version:

tamarind custom-tools test my-esmfold --version v3 \
  --input inputs.json --name esmfold-smoke
tamarind status esmfold-smoke
tamarind wait esmfold-smoke --timeout 600

Or use the SDK:

from tamarind import Tamarind

with Tamarind() as client:
    tool = client.custom_tools.get("my-esmfold")
    job = tool.test({"sequence": "MKT..."}, version="v3", name="esmfold-smoke")
    print(job.job_name, job.id, job.status)

The command returns after submission and the run appears in the tool's website Test history. This executes the tool and uses compute; it is not a dry run. Omit name / --name to generate a unique name. Pass the version number (such as v3) from build or versions; the SDK resolves its execution pin and protects against deleting and recreating the tool. The selected build must be complete. Design splitting is handled by the backend and returns the parent job receipt. Normal submit commands keep their existing behavior. If submission times out, check the job name in the error details before retrying; it may already be running.

For the complete CLI workflow, source layout, build monitoring, publishing, and rollback examples, see the Custom Tools CLI guide.

Output for agents

Every command emits JSON when stdout is not a TTY, or with --json. Result documents are written to stdout. Typed command, transport, and usage errors are written to stderr as {"error":{"type":...,"message":...,"exitCode":...,"detail":...}}; detail is omitted when unavailable. A domain verdict can deliberately combine a nonzero exit with a result document: invalid validate output remains on stdout with exit 5, and a terminal unsuccessful job remains on stdout with exit 9 so callers can inspect its status. Exit codes are stable: 0 ok, 1 generic failure, 2 usage, 3 auth, 4 not-found, 5 validation, 6 rate-limit, 7 timeout (a bounded wait elapsed while the remote job may still run), 8 budget/quota exhaustion, and 9 remote job failure.

For agent workflows, always put a deadline on blocking commands. submit --wait and results --wait accept --timeout; the standalone wait command remains the easiest way to reattach to a durable job name.

Starting in 0.2, results requires --download DIR by default and never prints a presigned URL implicitly. This intentional safety boundary is why the release uses a new pre-1.0 minor version. The explicit --show-url escape hatch returns a credential-bearing, short-lived URL; do not use it in agent, CI, or shared logs.

Not every tool ships a runnable example — schema <tool> --example exits non-zero (with a clear message) when one isn't available, so a > job.yaml redirect never silently produces an empty file. Destructive commands (delete, files delete) require --yes/-y when run non-interactively (piped or --json), so an agent never removes data without an explicit confirmation.

tamarind --json jobs | jq '.jobs[] | select(.JobStatus=="Running")'

--json, --profile, and the endpoint/auth overrides are global options, so place them before the command name.

Commands

Run tamarind <command> --help for full options.

Configuration

Every request says whether it came from the CLI or the SDK, and which AI agent (Claude Code, Codex, Claude Science, ...) is running it when that can be told from the environment. Tamarind uses this only to count usage by channel and platform.

Profiles (key + endpoints) are stored in ~/.tamarind/config.json. Use a profile to point at staging:

tamarind --profile staging --api-base https://staging.tamarind.bio/api/ auth login
TAMARIND_PROFILE=staging tamarind tools

Links

• CLI docs — the hosted quickstart and command reference
• MCP server — the hosted option for claude.ai, ChatGPT, and other MCP clients
• API key — get your key and review rate limits

Questions or need help? Contact us at info@tamarind.bio.

License

Licensed under the Apache License 2.0.