Allowly Python SDK

Async Python client for the Allowly runtime API.

Load hosted agent instructions

from allowly import Allowly

async with Allowly(api_key=ALLOWLY_RUNTIME_KEY) as client:
    instructions = await client.agent_instructions(agent_name="Support agent")

Pass the returned text to your agent. It contains the action, resource and field schema, descriptions, and exact policy reference for that configured name. In both hosted and copied modes, your agent turns free text into structured Check JSON; Allowly Check decides permission. Execute only after an allow. Refresh the instructions after policy changes. This method was added in SDK 0.9.0 and requires a runtime supporting GET /v1/agent-instructions; earlier SDKs do not gain it. It uses the runtime API key, rejects invalid text responses, and never uses an availability fallback.

Check shared permissions by agent name

After an operator explicitly finishes shared-policy setup, the runtime can resolve the authorization for a registered, self-reported agent name:

from allowly import Allowly

async with Allowly(api_key=ALLOWLY_RUNTIME_KEY) as client:
    result = await client.check_by_agent(
        expected_policy_ref={"policy_id": "support", "policy_version": 1, "definition_sha256": "a" * 64},
        agent_name="Support agent", actions=["ticket.reply"],
        resource="ticket:one", context={"external": True},
        idempotency_key="support-ticket-one",
    )
    authorization_id = result.authorization_id  # Evidence link, resolved by the server.
    if result.results["ticket.reply"].decision == "allow":
        await send_reply()

The name is a routing label, not verified identity. context.agent_label is reserved: the server records {name, source: "self_reported"}. Name routing rejects clients configured with an agent token or token supplier and never uses availability fallbacks, including configured fail_open actions. Stop on an error or deny; pause confirm/escalate for review. Save the original named request, including expected_policy_ref, context, resource, cost and session. After approval, call check_by_agent again with that same name and unchanged inputs, using a fresh idempotency key, and execute only a fresh allow. The runtime checks the current name assignment; the returned authorization ID links evidence and does not replace the named continuation flow. Existing check behavior and identity-bound authorizations are unchanged. The named method requires SDK 0.9.0 or newer and a runtime supporting /v1/agent-check. Copy the exact policy reference from the generated setup instructions. A policy change requires a refreshed prompt; a mismatched reference stops the request before policy evaluation.

Confirmation receipts

Confirmation responses expose receipt, a pending resolution envelope or None for an older runtime/replay. The decision is approved or not_approved; legacy outcomes are rejected by this fresh rebuild. Poll client.receipts.fetch_signed(response.receipt.receipt_id) when a receipt is present, then verify it with your configured workspace and trusted keys. The signature authenticates the recorded client report, not a named human's identity or approval. Resolution does not dispatch an action. A standalone Check integration re-checks with its original authorization; native Execute continues its saved operation as described below.

This OPA rebuild requires the published PyPI package allowly-receipt-format==5.1.0 and accepts only receipt wire 5. Run uv sync from this repository; no sibling verifier checkout is required. Native continuation requires a runtime that supports /v1/executions/{operation_id}/continue; upgrading this SDK does not deploy that runtime. Earlier SDK packages do not gain these methods.

Resolution webhook setup and verification

Use a setup/CLI credential to manage the workspace's one existing callback:

import os
from allowly import Allowly, verify_resolution_webhook

async with Allowly(api_key=os.environ["ALLOWLY_SETUP_KEY"]) as setup:
    configured = await setup.resolution_webhook.configure(
        "https://customer.example/allowly-resolution"
    )
    signing_secret = configured.signing_secret  # Store privately for the receiver.
    configuration = await setup.resolution_webhook.get()
    deliveries = await setup.resolution_webhook.deliveries()
    # rotated = await setup.resolution_webhook.rotate()
    # await setup.resolution_webhook.disable()

event = verify_resolution_webhook(
    raw_body, request_headers,
    signing_secret=os.environ["ALLOWLY_RESOLUTION_SIGNING_SECRET"],
    expected_workspace_id=os.environ["ALLOWLY_WORKSPACE_ID"],
)

Runtime keys cannot manage callbacks. Only configure/rotate return the signing secret; configuration reads omit it. Coordinate changes with the existing receiver: URL changes, re-enabling, and rotation cancel queued older-version events. The delivery list contains at most 20 summaries, without raw payloads or secrets. These methods use the existing API contract; no new endpoint is added.

Preserve the exact raw request bytes before JSON parsing and reject duplicate signature headers at your HTTP boundary. Verification authenticates the fixed HMAC-SHA256 Standard Webhooks profile with a 300-second attempt-timestamp tolerance and validates the configured workspace and bounded event fields. It does not prove a human's identity or verify the referenced receipts. Store event IDs durably before acknowledging; duplicate deliveries must not become duplicate business jobs.

Select the saved job by its trusted workspace, review kind/ID, and original source receipt ID. A null or mismatched source cannot select a native Execute job. Read current prompt status, then wake continue_http_execution with the unchanged original operation and arguments. A verified callback or an approved status is not dispatch permission; Continue validates current permission before one claim. Do not run an extra enforcing Check before native continuation. For standalone Check integrations, perform the existing fresh Check instead. Rejected, invalid, mismatched, or unavailable review never permits dispatch.

MCP middleware ships inside this SDK: pip install 'allowly[fastmcp]', then from allowly.mcp import AllowlyMCPMiddleware. In TypeScript, it lives in the separate @allowly/mcp package.

Author and preview an immutable policy

Action entries may include approved evidence requirements keyed by every declared field: AI {sources, instruction}, deterministic {source, transform: "email_domain", empty_sources}, or a boolean {all_of: [direct_boolean_field, ...]}. Requirements are retained in the whole definition digest. Omitted evidence stays omitted. Rules still use OR. SDK 0.9.0 uses the published verifier 5.1.0 for these policy and evidence validation helpers. No sibling checkout or source archive is required.

Policy create/get/list/archive and setup preview/publish require a setup/CLI credential, not a runtime API key or an agent token. Use the runtime client for Check and native Execute. The SDK does not change credentials or sign agent identity for policy setup calls.

from allowly import Allowly, PolicyDefinition

definition: PolicyDefinition = {
    "profile": "allowly.policy.v1",
    "agent_id": "research_agent",
    "actions": [{
        "policy": {
            "profile": "allowly.basic.v1", "action": "web.search",
            "decision": "allow", "target": None,
            "fields": {"risk": "integer"},
            "rules": [{"id": "review-high-risk", "decision": "confirm",
                       "target": "agent_owner",
                       "conditions": [{"field": "risk", "op": "gte", "value": 2}]}],
        },
        "guards": {}, "executable_operations": [],
    }],
    "default_expiry_days": 30,
}

async with Allowly(api_key=ALLOWLY_SETUP_KEY) as setup:
    created = await setup.policies.create(policy_id="research_agent", definition=definition)
    preview = await setup.policies.preview(definition=definition, action="web.search", context={"risk": 2})
    assert preview.provisional
    print(preview.policy_eval.result.decision, preview.policy_eval.input_sha256)
    retained = await setup.policies.get("research_agent")  # Latest active edition.
    page = await setup.policies.list(limit=20)
    # Publish a successor only when ready, naming the exact previous reference:
    # published = await setup.policies.publish(
    #     policy_id="research_agent", definition=updated_definition,
    #     previous_policy_ref=retained.policy_ref, idempotency_key="publish-research-2",
    # )
    # await setup.policies.archive("research_agent")

The complete definition and policy_ref keep their canonical snake_case JSON fields, ordered rules, and exact executable grants. Conditions use eq, ne, gt, gte, lt, lte, in, not_in, exists, contains_any, contains_none, or empty. A rule matches if any condition matches (OR). Valid comparisons use fixed priority: deny, escalate, confirm, then allow. Rules cannot weaken the action decision. Action deny always wins. Otherwise, a matching rule wins a tie with the action; the first matching rule wins ties between rules. If no rule matches, the action decision stays unchanged. List fields support overlap with 1–64 scalar values, or empty with a strict boolean value. Declared comparison types are strict; integers must be safe JSON integers, not floats or booleans. List items may mix strings, integers, booleans, and null, but not nested arrays or objects. Allow/deny targets are null, confirm targets are agent_owner, and escalation targets are actual member:<account_user_id> values. Publishing checks agent ownership, setup provenance, current membership, and executable catalog bindings on the server. A guessed member or team label is not a valid review target.

Preview calculates only provisional OPA rules. It does not check grants, live guards, quotas, budgets, or final execution permission. policy_eval.result can differ from Check's final decision. Parsing it or verifying a receipt signature is not independent policy replay; replay needs the original retained artifacts and the pinned official native OPA executable. The SDK has no parallel Python rule interpreter or unsigned fallback for malformed setup responses.

Subject authorization pattern

Do not send raw user/customer PII to Allowly receipts unless you intentionally want it in your audit trail. Create one authorization per subject, store the returned authorization ID in your own app database, and use that ID for later checks.

import asyncio
import os

from allowly import Allowly


async def main() -> None:
    async with Allowly(
        api_key=os.environ["ALLOWLY_API_KEY"],
        base_url=os.getenv("ALLOWLY_API_URL", "https://api.allowly.ai"),
    ) as allowly:
        # Your app creates a stable internal subject ID.
        subject_id = "subject_abc123"

        # Store this in your app table, for example:
        # allowly_authorizations(subject_id, policy_id, allowly_authorization_id, status)
        authorization = await allowly.authorizations.create(
            user_id=f"subject:{subject_id}",
            policy_id="research_agent",
            policy_version=1,  # The exact retained edition selected during setup.
            metadata={"source": "import"},
        )

        # Before the agent acts, check whether this action is allowed.
        decision = await allowly.check(
            authorization_id=authorization.authorization_id,
            actions=["web.search"],
            resource=f"subject:{subject_id}",
            context={"stage": "research", "risk": 1},
        )

    if decision.results["web.search"].decision != "allow":
        raise RuntimeError("Action is not authorized")


asyncio.run(main())

Allowly agent identity

Create the agent in the dashboard, then run allowly agent enroll <exact-agent-id> from the Allowly CLI. Store the resulting private credential on the trusted machine that runs the agent. Define its policy, then create a new authorization for that agent. CLI-only setups can still create a live policy before enrollment. The credential identifies the agent; the authorization and policy still decide what it may do. The workspace runtime API key is still required.

import os
from allowly import Allowly, NativeAgentCredential

credential = NativeAgentCredential.from_file("/secure/path/agent.json")
allowly = Allowly(
    api_key=os.environ["ALLOWLY_API_KEY"],
    agent_token_supplier=credential.token,
)

decision = await allowly.check(
    authorization_id="auth_...",
    actions=["order.submit"],
)

The SDK signs a fresh 60-second token for each request. The private key remains in your runtime; do not commit or log the credential file. The CLI registers only its public key with Allowly.

Existing Auth0 agent identity

For an authorization bound to an Auth0 machine identity, supply the short-lived access token separately from the Allowly runtime key. A token supplier runs for each check or execution. Use your existing OAuth client library to fetch and cache Auth0 tokens; keep the client secret outside this SDK. New self-service Auth0 setup is unavailable. Contact us to add your own identity provider.

from datetime import datetime, timezone

allowly = Allowly(
    api_key=os.environ["ALLOWLY_API_KEY"],
    agent_token_supplier=get_auth0_agent_token,
)

decision = await allowly.check(
    authorization_id="auth_...",
    actions=["order.submit"],
    client_timestamp=datetime.now(timezone.utc),
)

The SDK sends the agent token in the X-Allowly-Agent-Token header for checks and customer-hosted execution calls. Provider credentials stay in your runtime.

Receipt acknowledgments

After fetching the complete signed receipt, the included verifier can calculate the canonical receipt hash for an acknowledgment:

from allowly.verify import hash_seal_value, verify_receipt

claims = verify_receipt(signed_receipt, keys, expected_workspace_id=workspace_id)

ack = await allowly.acknowledge_receipt(
    receipt_id=claims["receipt_id"],
    receipt_sha256=hash_seal_value(signed_receipt),
    client_timestamp=datetime.now(timezone.utc),
    idempotency_key=f"ack:{claims['receipt_id']}",
)

# Retrieve the same acknowledgment later by its returned ID.
ack = await allowly.get_receipt_acknowledgment(
    claims["receipt_id"],
    ack.acknowledgment_id,
)

Client timestamps are customer-reported and must include a timezone. They do not replace the timestamp issued by Allowly in a receipt.

Execute from customer code

Enable an app in Settings → Executables, then bind its exact operation to the policy action. Enabling an app alone does not grant permission. The required receipt verifier is included with this SDK.

from allowly import Allowly

async with Allowly(api_key=ALLOWLY_RUNTIME_KEY, agent_token=AGENT_TOKEN) as allowly:
    result = await allowly.execute_http(
        "https://api.stripe.com/v1/refunds",
        operation_id="refund-order-123",  # stable across retries
        authorization_id="auth_...",
        enabled_executable_id="exe_...",
        catalog_operation_id="stripe.refunds.create",
        action="billing.refund",
        method="POST",
        headers={"authorization": f"Bearer {STRIPE_KEY}",
                 "content-type": "application/x-www-form-urlencoded"},
        body="payment_intent=pi_example&amount=2500",
        evidence_mode="receipt",
    )
    # result.response stays local; Allowly receives commitments and outcome metadata.

Use the operation ID from the catalog returned by your deployment; the initial catalog contains documentation-checked candidates, not a live compatibility guarantee. Policy inputs such as policy_input={"context": {"amount": 2500}} are customer-reported. A receipt does not prove that those inputs match the meaning of the provider body.

The helper calls remote /v1/execute, validates the approval, and claims dispatch once before sending locally. It fails closed on unavailable checks and stops on deny. Confirmation and escalation return status="waiting_for_review" with a typed review containing its kind, opaque ID, source receipt ID, and expiry. No provider request is sent while review is pending. Low-level customer methods are prepare_execution, continue_execution, claim_execution_dispatch, report_execution_outcome, and get_execution.

Read the review without running another Check or creating a receipt:

if result.execution.review.kind == "confirm":
    status = await allowly.confirmations.get_status(result.execution.review.id)
else:
    status = await allowly.escalations.get_status(result.execution.review.id)

These typed responses return the recorded choice, source/resolution receipt IDs, and current grant lifecycle. Use the opaque cnf_ monitor ID for confirmation, never its bearer nonce. status="approved" is only a wake-up signal: continue the saved operation so Allowly validates current permission. Null or unknown evidence is not an allow. await allowly.readiness() reads /readyz; readiness is not permission, and network/protocol errors raise rather than returning an allow.

Allowly receives the origin, path, query string and policy inputs. Header values and body bytes are committed by hash. Keep provider credentials in local headers, not in the URL, query string or policy context.

Keep the original URL, method, headers, body, and policy inputs in your own durable job store. A verified approval webhook can wake that job; polling the approval resource is another option. A webhook is not permission to send. Continue with the same arguments and operation ID:

result = await allowly.continue_http_execution(original_url, **original_arguments)

original_arguments must include the original operation_id, authorization, executable, action, headers, body, policy inputs, evidence mode, and journal directory. Restore credentials from your local credential store. The SDK checks the saved request hash and sends only the original request commitments and review binding to /v1/executions/{operation_id}/continue. The runtime validates the review and current permission before the SDK claims dispatch. Still-pending review remains waiting; rejected, expired, or revoked permission never sends. Do not call a separate enforcing Check before continuation: continuation itself performs the fresh evaluation, including any one-use escalation grant.

The private journal defaults to .allowly/executions; put it on persistent storage and use a separate directory per workspace. It retains the original prepare timestamp, request hash, review binding, and outcome, not raw provider headers, request body, or confirmation bearer nonce. The initial response can carry that nonce for your confirmation flow; do not log it. Version-1 journals and incomplete initial prepare attempts cannot be reopened through continuation. A repeated Execute raises ExecutionRecoveryRequired with its directory. Reconcile with get_execution or call flush_execution_outcome(operation_dir) to retry a saved report. Neither repeats the provider action. An interrupted or timed-out send may already have acted; never generate a replacement operation ID automatically. There is no exactly-once guarantee for arbitrary providers. If result.outcome_pending is true, its API response still describes approval; result.response contains the locally observed response and the journal retains the report for upload. The Python helper also raises ExecutionRecoveryRequired for a duplicate completed continuation instead of returning a cached local response. Use get_execution to read the outcome; do not turn that error into another send.

Signing stays asynchronous. To assemble the Allowly side of the evidence later:

from allowly import complete_execution_evidence

package = await complete_execution_evidence(
    allowly, result.operation_dir,
    public_keys=trusted_workspace_keys, expected_workspace_id="ws_...",
)

This verifies the decision receipt signature and its exact approval hash. The outcome remains customer-reported, not independent proof of execution or business completion.

FastMCP identity mapping

FastMCP middleware maps policy inputs explicitly. Raw tool arguments are available to the callback, but the middleware does not copy them into Allowly context. Provide trusted_user_id from your server's authenticated session or token.

from allowly.mcp import AllowlyMCPMiddleware, MCPCheckInput

middleware = AllowlyMCPMiddleware(
    api_key=os.environ["ALLOWLY_API_KEY"],
    user_id_fn=trusted_user_id,
    authorization_id_fn=lambda user_id: authorization_id_for(user_id),
    agent_token_fn=lambda request: auth0_agent_token_for(request.fastmcp_context),
    check_input_fn=lambda request: MCPCheckInput(
        action="email.send",
        resource=f"gmail:thread:{request.arguments['thread_id']}",
        context={"recipient_domain": request.arguments["recipient_domain"]},
        client_timestamp=datetime.now(timezone.utc),
        idempotency_key=request.arguments["operation_id"],
    ),
)
mcp.add_middleware(middleware)

The three resolver callbacks may be synchronous or asynchronous. When an agent_token_fn is configured, an error or an empty result blocks the tool before the check. Map only the fields the policy needs; ordinary tool arguments do not automatically satisfy policy context.

Local development against the documented Caddy endpoint requires the edge token that Cloudflare injects for public traffic. Pass it explicitly:

Allowly(
    api_key=os.environ["ALLOWLY_API_KEY"],
    base_url="http://localhost:8443",
    dangerously_allow_insecure_base_url=True,
    edge_token=os.environ["ALLOWLY_EDGE_TOKEN"],
)

The token is only sent when provided; never set it for the public API.

Send JSON through a private SEAL webhook

Copy the private URL from the dashboard's SEAL page. The URL is the only credential this client sends; it does not use an ordinary API key.

import asyncio
import os

from allowly import SealWebhookClient


async def seal_event(raw_json: str, event_id: str):
    async with SealWebhookClient(os.environ["ALLOWLY_SEAL_WEBHOOK_URL"]) as webhook:
        delivery = await webhook.send(
            raw_json,
            idempotency_key=event_id,
            type="invoice",
            reference="INV-1042",
            statement="Approved for payment",
        )
        while delivery.status in {"received", "signing"}:
            await asyncio.sleep(1)
            delivery = await webhook.get_delivery(delivery.attempt_id)
        if delivery.status != "sealed":
            raise RuntimeError(delivery.error_code or "SEAL delivery failed")
        return delivery.receipt, await webhook.get_keys()

The webhook processes your JSON to create a fingerprint; Allowly stores the fingerprint and signed receipt. Keep the original record in your workflow. Receipt details are sent in the three explicit Allowly-Seal-* headers. Their values must use printable ASCII, may contain interior spaces, and must not have leading or trailing whitespace. The client rejects invalid values instead of changing them. Direct API metadata still supports its existing Unicode values. When a signed receipt is present, the client returns its signed metadata and rejects a conflicting top-level delivery projection. Treat the full URL like a password and keep it out of logs, tickets, and source control. Regenerating or disabling it stops the old URL. Delivery associations and status remain available for 7 days; preserve signed receipts and keys under your own retention policy. With no idempotency_key, retrying after a lost response can create another seal.

Seal a JSON record with local hashing

seal hashes strict raw JSON in your process, sends only its digest to Allowly, and waits for the full signed receipt. Generate and persist request_id in your workflow so a retry recovers the same seal:

import uuid

request_id = str(uuid.uuid4())
sealed = await allowly.seal(
    raw_json,
    request_id=request_id,
    metadata={"source": "invoice-workflow"},
)
save_beside_record(sealed.receipt)

Use seal_value(parsed_json, ...) only when the original JSON text is no longer available. A parsed value cannot reveal duplicate object names or the original number spelling, so seal is the safer input boundary.

To verify later, preserve the authenticated workspace_id response and a key document fetched from Allowly through an authenticated or previously trusted source. Keep the signature and record checks separate:

from allowly.verify import load_keys_from_json, verify_seal_json

result = verify_seal_json(
    raw_json,
    sealed.receipt,
    load_keys_from_json(keys_doc),
    expected_workspace_id=sealed.workspace_id,
    trusted_key_fingerprints=configured_key_fingerprints,
)
assert result.signature_verified
assert result.record_matches

Authorization creation requires a logical policy_id and exact positive policy_version. Optional actions is a names-only subset of that edition; omit it to select all its actions. Inline policy rules, action objects, agent_id, constraints, and decision-override maps are not accepted.

Unavailable checks fail closed unless an action is explicitly mapped to "fail_open" with fallback_by_action. Unmapped actions always fail closed. Identity-enabled checks always fail closed, including token supplier failures and identity_verification_unavailable responses.

For actions that need third-party approval, define the escalation rule on the agent policy, create the authorization from its exact policy_id and policy_version, and then resolve returned escalation results with await allowly.escalations.approve(escalation_id, resolved_by="manager:123") or reject(...), then re-check before running the action.

Read a confirmation or escalation without resolving it:

# confirmation_id comes from a confirm check result. It is not confirm_nonce.
confirmation = await allowly.confirmations.get(confirmation_id)
escalation = await allowly.escalations.get(escalation_id)
print(confirmation.status, confirmation.authority_status)
print(escalation.status, escalation.authority_status)

The released get names remain aliases of get_status; both return the same typed status object. ConfirmationStatusResponse and EscalationStatusResponse are aliases of the existing ConfirmationStatus and EscalationStatus types. Each read retries once if the server disconnects without a response. A second failure or malformed response raises an error; keep the workflow paused. These GET retries never replay a Check or other write. Repeat reads in your application's own bounded polling loop if needed. The prompt status is pending, approved, rejected, expired, or unknown; unknown means a legacy record does not show the choice. An approved choice stays approved after its grant expires, is revoked, or (for escalations) is consumed. authority_status="available" is a lifecycle snapshot, not permission. For standalone Check integrations, make a fresh allowly.check(...) with the original authorization and execute only an allow. For native Execute, continue the saved original operation; do not insert an extra enforcing Check. These reads never execute actions, consume approval, or create receipts. Nullable receipt IDs refer to existing records. Older check responses may omit confirmation_id.

If you need lookup by email later, import from_email from allowly.identifiers and store from_email(email, pepper=APP_PII_PEPPER). The helper trims and lowercases only, prefixes the result with email_hmac:v1, and never sends the raw email or pepper to Allowly. Keep the pepper stable and backed up; changing it changes derived user IDs. Keep raw names, emails, documents, and profile URLs out of Allowly receipts unless those fields are intentionally part of your audit record.

Do not add raw HTTP fallbacks in application code for APIs the SDK is missing. Patch this SDK first, then use the typed client from the app. That keeps the integration examples honest and makes SDK gaps visible early.

Offline receipt verification

Use the included verifier to hash SEAL records and verify signed receipts. This SDK pins the published allowly-receipt-format==5.1.0, which verifies receipt wire format 5 (the package major equals the wire format). alg and the key ID are in the signed JWS protected header. The receipt has only protected, payload, and signature fields, all base64url strings. Store and share that original envelope. verify_receipt returns verified claims for decisions and request binding. decode_receipt only reads claims for display; it does not prove their signature or grant permission.

Wire format 5 preserves daily receipt.checkpoint commitments. For an untrusted receipt bundle, pass caller-trusted key fingerprints to verify_receipt; a fingerprint copied from that same bundle is not a trust anchor.

Key-document fetching requires HTTPS by default. For the documented local Caddy endpoint only, pass dangerously_allow_insecure_base_url=True and its edge_token to fetch_keys_doc, matching the client options above.