Connect your Python application

Use the SDK when your application needs to submit observations, check their review state and retrieve knowledge. You need a running gate's MCP endpoint and whatever credential that deployment requires. Ask its operator for the endpoint and access; a URL or space name alone does not grant permission.

1. Install the development SDK

From a checkout of the SDK revision you intend to use:

python -m venv .venv
. .venv/bin/activate
python -m pip install .

Development distributions are published through GitHub, not PyPI. These commands install the selected source checkout; they do not start a gate or provision access.

2. Connect and submit an observation

Set DOUBLEGATE_ENDPOINT to the full MCP URL. Supply DOUBLEGATE_TOKEN through your application's secret mechanism when the deployment requires a bearer credential. The HTTPS address below must be your actual deployment, not an example host.

import os
from doublegate_sdk import connect

client = connect(
    os.environ['DOUBLEGATE_ENDPOINT'],
    token=os.environ.get('DOUBLEGATE_TOKEN'),
    allow_writes=True,
)
proposal = client.propose(
    'The laboratory labels basalt specimens by collection date.',
    content_type='memory',
    trust_class='T-4',
    source_uri='application://laboratory/observation/123',
)
artifact_id = proposal['artifact_id']
print(client.status(artifact_id)['state'])

Keep artifact_id: it is the handle for checking this proposal. The returned state is the gate's answer. Submission success does not mean review is finished.

3. Retrieve knowledge

answers = client.recall('How are basalt specimens labelled?', limit=5)
for result in answers['results']:
    print(result)

An empty result immediately after submission is expected while the observation is awaiting review. Recall excludes your unreviewed submissions. Provisional knowledge is excluded by default; use include_provisional=True only when your application accepts it. Keep the returned provenance and flags with the content you use.

Run the complete example

The repository includes the same workflow as a runnable program:

python examples/connect_to_gate.py --endpoint "$DOUBLEGATE_ENDPOINT"

This writes one observation to the selected gate. It never launches a substitute server. For a local development endpoint, explicitly add --allow-insecure-loopback; remote connections use HTTPS.

Handle failures

Catch GateError from doublegate_sdk.client and branch on kind, not its text. If outcome_unknown is true after a write, reconcile with the gate before retrying: the server may already have accepted the request. The SDK does not retry writes. See client reference for the exact errors and bounds.

Deployment boundary

Use the MCP URL, not the daemon administration /rpc URL. Both use HTTP but have different operations. See HTTP interfaces and migration.

The inspected Client Gate's local /mcp endpoint is keyless on loopback. Its console token protects /api/*, not that MCP route. An SDK bearer header cannot supply missing server enforcement. Authenticated remote MCP requires an endpoint that actually verifies the credential. See roadmap for that acceptance work; do not expose the keyless local listener as a remote service.

Next: connect an agent host, or write a deterministic check if you are extending validation rather than consuming memory.