Tilde Python Sdk

Tilde Python SDK

Python SDK for the Tilde data versioning API.

Installation

pip install tilde-sdk

Or with uv:

uv add tilde-sdk

Requires Python 3.11+.

Quick Start

import tilde

repo = tilde.repository("my-org/my-repo")

# Run commands in an interactive sandbox
with repo.shell(image="python:3.12") as sh:
    sh.run("pip install pandas")
    result = sh.run("python train.py")
    print(result.stdout.text())

    # Stream output line by line
    result = sh.run("cat /sandbox/results.csv")
    for line in result.stdout.iter_lines():
        print(line)

> [!IMPORTANT] > Transactional by default. All filesystem modifications made in a sandbox > happen in the context of a transactional session. If anything fails midway or > is aborted, changes don't take effect. Only successful sandboxes' changes are > committed atomically to the repository -- so your data is always in a > consistent state. See [Sessions](#sessions) for more details.

Authentication

The SDK resolves credentials in this order (first match wins):

1. Explicit parameter — Client(api_key=...) or tilde.configure(api_key=...).
2. Environment variable — TILDE_API_KEY.
3. CLI config file — ~/.tilde/config.yaml (written by tilde auth login).
4. Sandbox metadata endpoint — TILDE_SANDBOX_CREDENTIALS_URI (auto-detected

only when no static key is configured; injected by the sandbox runtime for code running inside a Tilde sandbox).

A static key always wins over sandbox auto-detection, so credentials set deliberately by the caller are never silently overridden.

Using the CLI's saved credentials

After tilde auth login, the SDK picks up the credentials automatically — no env var or in-code configuration needed:

import tilde

repo = tilde.repository("my-org/my-repo")

Environment variables

Right for CI/CD, agent workflows, and Docker containers where the CLI hasn't run:

export TILDE_API_KEY="your-api-key"
export TILDE_ENDPOINT_URL="https://tilde.run"  # optional

Module-level configuration

import tilde

tilde.configure(api_key="your-api-key")
repo = tilde.repository("my-org/my-repo")

Explicit client

Use an explicit client when you need multiple configurations or full control over the HTTP lifecycle:

from tilde import Client

with Client(api_key="your-api-key") as client:
    repo = client.repository("my-org/my-repo")

A missing API key is not an error at construction time; a ConfigurationError is raised when the first request is made.

Configuration reference

| Option | Environment variable | Default | |--------|---------------------|---------| | api_key | TILDE_API_KEY | required | | endpoint_url | TILDE_ENDPOINT_URL | https://tilde.run | | default_sandbox_image | TILDE_DEFAULT_SANDBOX_IMAGE | ubuntu:22.04 |

Sandbox Execution

The fastest way to run code against your repository data. Sandboxes execute inside isolated containers with your data mounted as a volume — every change is captured as a transaction.

Interactive shell

Use repo.shell() to run multiple commands in a single sandbox:

with repo.shell(image="python:3.12") as sh:
    sh.run("pip install pandas")
    result = sh.run("python train.py")
    print(result.stdout.text())
    print(result.stderr.text())

The shell is a context manager — on clean exit changes commit automatically, on exception they roll back. Each sh.run() spawns an independent child process: shell state (cd, exported vars, aliases) does not persist across calls. Chain commands in one call (sh.run("cd /foo && ls")) or pass env= / cwd= to run().

shell.run() returns a RunResult with stdout, stderr, and exit_code. Pass check=True to raise CommandError on non-zero exits.

One-shot execution

For a single command that doesn't need an interactive session:

result = repo.execute("python train.py", image="python:3.12")
print(result.stdout.text())

# check=False to handle errors yourself
result = repo.execute("might-fail", image="python:3.12", check=False)
if result.exit_code != 0:
    print(result.stdout.text())

Output streams

RunResult.stdout and RunResult.stderr are OutputStream instances:

| Method | Returns | Description | |--------|---------|-------------| | .read() | bytes | Full output as raw bytes | | .text(encoding='utf-8') | str | Full output decoded as a string | | .iter_bytes(chunk_size) | Iterator[bytes] | Yield byte chunks | | .iter_text(chunk_size) | Iterator[str] | Yield text chunks | | .iter_lines() | Iterator[str] | Yield lines (no trailing newlines) |

Repositories

import tilde

# Shorthand: the only top-level shortcut in the SDK
repo = tilde.repository("my-org/my-repo")

# Lazy-loaded properties
print(repo.id, repo.description, repo.visibility)

# Update
repo.update(description="New description", visibility="public")

# Delete (soft delete)
repo.delete()

Create a repository through its organization:

org = tilde.organizations.get("my-org")
repo = org.repositories.create(
    "my-repo",
    description="My dataset",
    session_max_duration_days=7,
    retention_days=90,
)

for r in org.repositories.list():
    print(r.name, r.description)

Commits

# Newest first, auto-paginating
for commit in repo.commits.list():
    print(commit.id, commit.committer, commit.message)

# Cap results
for commit in repo.commits.list(amount=10):
    print(commit.id)

# Look up by ID
commit = repo.commits.get("a1b2c3d4e5f6")

# Diff introduced by a commit
for change in commit.diff():
    print(change.status, change.path)

# Revert
revert = commit.revert(message="undo")

Pagination

Every .list() returns a PaginatedIterator that fetches pages lazily. Two keyword arguments tune iteration:

| Argument | Effect | |---|---| | amount | Cap on the total number of results yielded. | | page_size | Number of results per HTTP page (default 100, server max 1000). |

for entry in commit.objects.list(prefix="data/", page_size=500):
    process(entry)

Sessions

Sessions provide direct transactional access to objects. Prefer sandbox execution for running code; sessions are for fine-grained object operations from your own process.

# Context manager — rolls back on error
with repo.session() as session:
    session.objects.put("data/report.csv", b"content")
    session.objects.delete("data/old.csv")
    session.commit("update CSV files")

# Explicit control
session = repo.session()
print(session.session_id)
session.objects.put("data/file.csv", b"content")
session.commit("modifying data")
# or: session.rollback()

Resume a session from another thread, process, or machine:

session = repo.attach(session_id)
session.objects.put("data/file2.csv", b"more content")
session.commit("finishing work")

Objects

Writing

import pathlib

with repo.session() as session:
    # From bytes
    session.objects.put("data/hello.txt", b"Hello, Tilde!")

    # From a file
    with open("dataset.parquet", "rb") as f:
        session.objects.put("data/dataset.parquet", f)

    # From a Path (opened/closed automatically)
    session.objects.put("data/model.bin", pathlib.Path("model.bin"))

    # Server-side copy (no download/upload)
    session.objects.copy("data/hello.txt", "data/hello-backup.txt")

    # Delete
    session.objects.delete("data/old.csv")
    session.objects.delete_many(["data/a.csv", "data/b.csv"])

    session.commit("upload files")

Files ≥ 64 MB automatically use multipart upload; smaller files use a single presigned PUT. No code changes needed.

Reading

Objects can be read from a committed snapshot (via a Commit) or from within a session (including uncommitted staged changes):

# From a commit
commit = next(iter(repo.commits.list()))
with commit.objects.get("data/hello.txt") as f:
    print(f.read().decode())

# Within a session
with repo.session() as session:
    with session.objects.get("data/file.csv") as f:
        data = f.read()

    # Metadata only
    meta = session.objects.head("data/file.csv")
    print(meta.etag, meta.content_type, meta.content_length)

Streaming and byte ranges

Large objects: disable caching and stream in chunks:

with commit.objects.get("data/large.bin", cache=False) as f:
    for chunk in f.iter_bytes(chunk_size=1024 * 1024):
        output.write(chunk)

Byte ranges work for file headers, log tails, or columnar formats like Parquet:

# First 4 bytes
with commit.objects.get("data/file.parquet", byte_range=(0, 3)) as f:
    magic = f.read()

# From offset 1024 to end
with commit.objects.get("data/file.parquet", byte_range=(1024, None)) as f:
    tail = f.read()
    print(f.content_range)   # "bytes 1024-49151/49152"
    print(f.content_length)  # bytes returned

Listing

for entry in commit.objects.list(prefix="data/", delimiter="/"):
    print(entry.path, entry.type)  # "object" or "prefix"

Uncommitted changes

session = repo.session()
session.objects.put("data/new.csv", b"content")
for entry in session.uncommitted():
    print(entry.path)

Organizations

import tilde

# Create, list, get, delete at the module level
org = tilde.organizations.create("my-org", display_name="My Org")
for o in tilde.organizations.list():
    print(o.name, o.display_name)
org = tilde.organizations.get("my-org")
tilde.organizations.delete("my-org")

An Organization exposes every org-scoped sub-resource as a property:

org = tilde.organizations.get("my-org")
org.repositories
org.members
org.agents
org.roles
org.groups
org.policies
org.connectors

Members

for m in org.members.list():
    print(m.username, m.email)

org.members.create("alice")          # add by username
org.members.delete("user-uuid")      # remove by user_id

Groups

groups = org.groups

group = groups.create("engineers", description="Engineering team")
group = groups.get(group.id)                                # includes members + attachments
group.members.create(subject_type="user", subject_id="user-uuid")
group.members.delete(subject_type="user", subject_id="user-uuid")
group.update(name="engineering", description="Updated")
group.delete()

# Effective groups for a principal
for g in groups.effective(principal_type="user", principal_id="user-uuid"):
    print(g.group_name, g.source)

Policies

policies = org.policies

# Validate before creating
result = policies.validate("GetObject()")
print(result.valid, result.errors)

policy = policies.create(
    name="read-only",
    policy_text="ListRepositories()\nGetRepository()\nGetObject()\n",
    description="Read-only access",
)

# Generate from natural language
text = policies.generate("Allow read-only access to all repositories")

# Attach / detach
policy.attach(principal_type="group", principal_id="group-uuid")
policy.detach(principal_type="group", principal_id="group-uuid")

# Effective policies for a principal
for ep in policies.effective(principal_type="user", principal_id="user-uuid"):
    print(ep.policy_name, ep.source)

Roles and Agents

Non-human identities authenticated via API keys:

• Roles (trk- prefix) — CI/CD pipelines, automation without metadata.
• Agents (tak- prefix) — automated tools and AI assistants, with metadata

and optional inline policies.

# Roles
role = org.roles.create("ci-deployer", description="CI/CD pipeline")
for role in org.roles.list():
    print(role.name)

# Agents
agent = org.agents.create(
    "data-pipeline",
    description="Nightly pipeline",
    metadata={"env": "prod"},
)
agent.update(metadata={"env": "staging"})
agent.update(inline_policy="ListRepositories()\nGetObject()\n")

API keys

agent = org.agents.get("data-pipeline")

# Create a key (the full token is only shown once)
created = agent.api_keys.create("pipeline-key")
print(created.token)              # tak-...

for key in agent.api_keys.list():
    print(key.name, key.token_hint)

key = agent.api_keys.get(key_id)
key.revoke()

The same api_keys collection exists on Role instances (tokens prefixed trk-).

Secrets

Encrypted key-value pairs injected as environment variables into sandboxes.

# Repository-scoped
repo.secrets.create("API_KEY", "sk-abc123...")
secret = repo.secrets.get("API_KEY")
print(secret.value)
repo.secrets.delete("API_KEY")

# Agent-scoped (override repo secrets with the same key)
agent = org.agents.get("data-pipeline")
agent.secrets.create("OPENAI_KEY", "sk-abc123...")

Precedence at sandbox launch (highest to lowest):

1. env= passed in the sandbox request
2. Agent secrets (if running as an agent)
3. Repository secrets

Connectors and Imports

connectors = org.connectors

# S3
conn = connectors.create(
    name="production-s3",
    type="s3",
    source_uri="s3://my-bucket/datasets/",
    config={
        "access_key_id": "AKIA...",
        "secret_access_key": "...",
        "region": "us-west-2",
    },
)

# Attach to a repo
repo.connectors.attach(conn.id)
for c in repo.connectors.list():
    print(c.name, c.type)

Import from a connector:

import time

job = repo.imports.create_from_connector(
    connector_id=conn.id,
    destination_path="imported/",
    source_prefix="datasets/",
    commit_message="Import production datasets",
)

while job.status not in ("completed", "failed"):
    time.sleep(2)
    job.refresh()
    print(job.status, job.objects_imported)

if job.status == "completed":
    print(f"Import done! Commit: {job.commit_id}")

Cross-repository imports copy data from one Tilde repo to another:

job = repo.imports.create_from_repository(
    repo_path="other-org/source-data",
    destination_path="external/",
    source_prefix="datasets/train/",
    commit_message="Import training data",
)

Agent Approval Workflow

When a repository requires approval for agent commits, session.commit() blocks by default until a human approves or rolls back. The approval URL is emitted as a UserWarning.

with repo.session() as session:
    session.objects.put("data/results.csv", b"col1,col2\na,b\n")
    session.commit("add results")  # blocks until approved

Non-blocking:

result = session.commit("add results", block_for_approval=False)
# result is None; session stays open for review

Structured result (no blocking, no warnings):

result = session.commit_result("add results")
if result.status == "committed":
    print(result.commit_id)
elif result.status == "approval_required":
    print(result.web_url)

Low-Level Sandboxes and Triggers

repo.shell() and repo.execute() cover most needs. For direct control over async lifecycle, triggers, and delegation, use repo.sandboxes and repo.sandbox_triggers. See the full docs for details.

Error Handling

All SDK exceptions inherit from TildeError:

TildeError                           # base for all SDK errors
├── ConfigurationError               # missing API key, bad endpoint
├── TransportError                   # network failures, DNS, timeouts
├── SerializationError               # invalid JSON in response
├── SandboxError                     # sandbox lifecycle failure
├── CommandError                     # non-zero exit (repo.execute / shell.run(check=True))
└── APIError                         # base for HTTP API errors
    ├── BadRequestError              # 400
    ├── AuthenticationError          # 401
    ├── ForbiddenError               # 403
    ├── NotFoundError                # 404
    ├── ConflictError                # 409
    ├── GoneError                    # 410

…

## Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [tilderun](https://github.com/tilderun)
- **Source:** [tilderun/tilde-python-sdk](https://github.com/tilderun/tilde-python-sdk)
- **License:** Apache-2.0

Install and usage instructions live in the source repository linked above.