annotide

Python SDK and command line for the Annotide API (API-3). One runtime dependency (httpx); responses are plain dicts typed with TypedDicts generated from the API's OpenAPI document.

pip install annotide          # or `pip install -e sdk` from a checkout
export ANNOTIDE_URL=https://annotate.example.com
export ANNOTIDE_API_KEY=... # Settings → API keys; use a service account for CI

Python

from annotide import Client, JobFailedError

with Client() as client:  # reads ANNOTIDE_URL / ANNOTIDE_API_KEY
    project = next(p for p in client.list_projects() if p["name"] == "Cars")

    # Freeze the dataset, then export it reproducibly.
    snapshot = client.take_snapshot(
        project["id"],
        "2026-q3",
        split={"train": 0.8, "val": 0.1, "test": 0.1, "seed": 0, "group_by": "folder"},
    )
    client.export(project["id"], "coco", "train.zip", snapshot_id=snapshot["id"], split="train")

    # Import existing labels; a dry run reports what would happen.
    report = client.import_file(project["id"], "labels.json", "coco", dry_run=True)
    print(report["result"])

• Errors are ApiError (status, title, detail from the RFC 9457 body); a job that ends failed/cancelled raises JobFailedError, one that outlives its timeout raises JobTimeoutError.
• A 429 waits for Retry-After and repeats. Gateway errors (502/503/504) repeat only when that is harmless: reads, and creates, which always carry an Idempotency-Key so a repeat answers the first row.
• Export downloads go straight to storage with the signed URL; the API key is never sent there.
• client.request(method, path, ...) reaches any endpoint without a method.

Command line

annotide whoami
annotide projects list | jq -r '.name'
annotide snapshots create <project> 2026-q3 --split 0.8,0.1,0.1 --group-by folder
annotide export <project> --format coco --snapshot <id> --split train -o train.zip
annotide import <project> labels.json --format coco --class-map '{"auto": "car"}' --dry-run
annotide jobs wait <job>

Lists print as JSON Lines, single objects as JSON; errors go to stderr with exit status 1. The key is read from ANNOTIDE_API_KEY only, so it never lands in shell history.

MCP server for AI agents (API-8)

annotide mcp serves the Model Context Protocol over stdio, so an AI agent (Claude Code, Claude Desktop, or any MCP client) can pre-label a project. It calls the REST API with an API key, so the agent can do exactly what that account may do. It can list projects and their label schema, view items (images come back as images), read annotations, claim and release tasks, and post pre-labels. It cannot submit, review, delete or export.

Setup, once per agent:

1. Service account. Settings → API keys → Service accounts: create one, mint a write key, and add the account to the project as an annotator.
2. Model. Models → Register model, name it after the agent, and leave the endpoint empty. That makes it an external producer: the platform never calls it, and the agent posts the pre-labels. Add version 1 and note its id.
3. Client config. For Claude Code (.mcp.json) or Claude Desktop:

{
  "mcpServers": {
    "annotation": {
      "command": "annotation",
      "args": ["mcp"],
      "env": {
        "ANNOTIDE_URL": "https://annotate.example.com",
        "ANNOTIDE_API_KEY": "ak_…",
        "ANNOTIDE_MODEL_VERSION_ID": "<model version id>"
      }
    }
  }
}

Install with pip install "annotide[mcp]". Pre-labels arrive as draft versions authored by the model version. People review and submit them as usual, and a pre-label never replaces human work (ML-10).

Development

make install                  # from the repository root: creates sdk/.venv
cd sdk && .venv/bin/pytest && .venv/bin/mypy annotide tests tools
make openapi                  # after an API change: docs/openapi.json + models.py

annotide/models.py is generated; never edit it by hand. CI fails when it or docs/openapi.json is stale.