Provenza

Provenza Python SDK

Governed, signed and auditable actions for AI agents.

pip install provenza

from provenza import Agent
agent = Agent("support-bot")   # reads PROVENZA_URL and PROVENZA_TOKEN
@agent.tool(name="zendesk.read")def open_tickets(status: str = "open") -> list:    return zendesk.tickets.list(status=status)
open_tickets()   # checked by Provenza, run, recorded and signed

What happens on a call

In connected mode (when PROVENZA_URL is set):

1. The SDK asks the server whether this call may run. The server checks the agent's rights, your organization's policies, the Tool Registry, behaviour, and the circuit breaker.

2. The tool runs only if the call is allowed. Otherwise:

   • a denied call raises ProvenzaDenied (a PermissionError);
   • a call that needs a human decision raises ProvenzaApprovalRequired.

   In both cases the attempt is recorded. Once a person approves the held call, make exactly the same call again: it runs once, within an hour.

   python

   try:    send_report(to="[email protected]")except ProvenzaApprovalRequired as held:    if agent.wait_for_approval(held.action_id)["decision"] == "approved":        send_report(to="[email protected]")      # the same call: runs once

3. The result is recorded and signed with the agent's own key. Provenza seals it into its signed audit log, so you can download a receipt for it that is verifiable offline.

The SDK fails closed: if the server cannot be reached, the tool does not run and the SDK raises ProvenzaUnavailable. Pass fail_open=True to run anyway and log the call locally as unchecked.

In standalone mode (no server), calls run and are written to a local, signed, hash-chained audit log. You can restrict them with allowed_tools=[...].

Only hashes of arguments and results are written to the local audit log, never the content.

Getting an identity

An admin issues a one-time enrollment token in Provenza (Agents → New agent). The agent's keys are generated on its own machine, and only the public key goes to the server:

export PROVENZA_URL=https://provenza.example.comexport PROVENZA_TOKEN=pvz_enr_...      # used once; afterwards the identity file is enough

The identity is stored in ~/.config/provenza/agents/<name>.json (mode 600). The SDK refuses to load it if other users can read it. It is the same file tools/provenza_sign.py enroll writes, so you can enroll with the CLI and run with the SDK.

To keep the private keys in the OS keychain instead:

pip install 'provenza[keychain]'

Agent("bot", key_store="keychain")

For post-quantum hybrid keys (Ed25519 + ML-DSA-65):

pip install 'provenza[pq]'

Agent("bot", hybrid=True)

agent.fingerprint is the value to give whoever verifies the agent's signatures.

More

with agent.chain(chain_id):            # act inside a delegation chain    open_tickets()
agent.delegate(other_agent_id, "summarize Q3", ["read"], tools=["db.read"], expires_in=600)
chk = agent.check("files.delete", {"path": p})   # ask for something that is not a decorated toolif chk.allowed:    delete(p)    agent.record(chk, {"deleted": True})

Async tools work the same way (async def). If a call ran but could not be recorded on the server (the server went down mid-call), the SDK does not hide the result. It writes a signed gap entry to the local audit log and emits a ProvenzaWarning.

Latency

A governed call costs the SDK's own work plus round-trips to the server. The SDK's own work was measured against an in-process server (no network) on 2026-10-11:

Add your network round-trip (×2 or ×1) and the server's checks.

With "sync", the record is on the server when the tool returns.

With "background", records are sent from a queue, in order. A failure becomes a signed gap in the local audit log, the same as in sync mode. flush() / aflush() wait for the queue, and close() and interpreter exit drain it. A process that is killed loses whatever is still queued; that is the price of not waiting.

Refusals are always recorded before the exception is raised, so ProvenzaDenied.action_id is known at once.

Auto-instrumentation

import provenza.autoprovenza.auto.instrument(agent)        # httpx (sync and async) and requests

Every direct call from this process to an AI provider becomes a governed action. That covers the OpenAI, Anthropic, Mistral, Cohere and Gemini SDKs, LangChain's provider clients, or plain HTTP. The provider hosts come from your organization's AI domain catalog (Shadow AI), and the server decides each call:

Only the host, method, path (without the query string), and the hash and size of the body are recorded. The prompt is never recorded.

A refusal comes back as an HTTP answer: 403 with an OpenAI-style error body and x-should-retry: false. Provider SDKs therefore raise their usual permission error with Provenza's reason and do not retry. With the server unreachable, nothing is sent (503). Other hosts, and calls to Provenza itself, are untouched. provenza.auto.uninstrument() puts everything back.

Egress proxy

Behind a Provenza egress proxy (proxy/ in the repository) the agent has two options. It can send its key and let the proxy check every request, or it can check a request itself and send the check id:

body = json.dumps({"amount": 1000}).encode()chk = agent.proxy_check("stripe", "POST", "/v1/charges", body=body)if chk.allowed:    httpx.post("http://127.0.0.1:8443/stripe/v1/charges", content=body, headers=chk.proxy_headers)

path, query and body must be exactly what goes on the wire. The proxy refuses any other request under that check. The proxy records the request, so it is not passed to record().

agent.proxy_request_headers(...) (same arguments) returns the headers to send: the check's, or, while Provenza is unreachable, only the agent's signed request. A proxy that enforces the signed policy then judges it by the signed release alone until its local kill switch stops it, and Provenza records the request when it is back (docs/pvp/edge.md, section 6).

HTTP tools through Provenza

An API registered in Provenza (Gateways → HTTP tools) is called through Provenza, and the agent never holds its key. The call is checked, made and recorded in one step. A tool that takes a credential is called through the egress proxy, which holds it - Provenza itself holds no credential:

agent = Agent("billing-bot", egress_proxy="http://127.0.0.1:8443")   # or PROVENZA_EGRESS_PROXYout = agent.invoke("crm.set_tier", {"id": "42", "tier": "gold"})out["status"], out["body"]          # a refusal raises ProvenzaDeniedagent.http_tools()                  # names, descriptions and input schemas, for the model

Through a proxy the HTTP request itself is signed too (X-Provenza-Request), which a proxy in enforce mode requires on every path. For clients that cannot sign (a stock OpenAI or MCP client) run the proxy's agent mode next to the agent with this identity file (provenza-proxy -agent-identity …, proxy/README.md).

Egress self-check

Is this host confined to the Provenza egress proxy, its only way out? Run the check at start-up and then periodically. With Require confined egress on, an agent acts only while its latest check is confined. See docs/egress-lockdown.md.

agent.egress_selfcheck()     # {"verdict": "confined" | "open", ...}

Permits for tools

An allowed check carries a permit signed by the organization. The agent sends it with the call:

chk = agent.check("crm.update", payload)if chk.allowed:    httpx.post("https://crm-tool.internal/update", json=payload, headers=chk.permit_headers)    agent.record(chk, output)

The tool runs nothing without a valid permit for exactly this call:

from provenza.permit import PermitMiddleware, PermitVerifier, require_permit
verifier = PermitVerifier("https://provenza.example.com", org="acme")app = PermitMiddleware(app, verifier, tool="crm.update")   # any ASGI app: FastAPI, Starlette, ...
@require_permit(verifier, "db.read")                         # or a plain functiondef read(query: str): ...

The verifier checks the signature against the organization's JWKS, the tool, the hash of the JSON body (it must be the checked input) and the expiry. It then redeems the permit with Provenza, so it is used once across every replica of the tool. If Provenza cannot be reached, it refuses.

Integrations

Each framework is an optional extra. The SDK never requires one: crewai and autogen-core even need conflicting protobuf versions.

Through Provenza's gateways. The call goes through the server, which holds the provider credentials, so it is enforced whatever the agent's code does.

# pip install 'provenza[openai]'from provenza.integrations.openai import gateway_client, wrapclient = gateway_client(agent)              # an openai.OpenAI pointed at the LLM gateway, with the agent's keyclient = wrap(existing_client, agent)       # or a copy of one you already configured
# MCP: no extra needed. Point any MCP client at Provenza instead of at the servers.from provenza.integrations.mcp import connection, client_configconn = connection(agent)                    # {"url": ".../api/v1/mcp", "headers": {...}}

MCP clients that start a local command (Claude Desktop, Claude Code, Cursor, Windsurf, Gemini CLI, VS Code): provenza-mcp is the same gateway over stdio. Issue an enrollment token in Provenza (Agents → New agent) and add:

{"mcpServers": {"provenza": {  "command": "uvx", "args": ["provenza-mcp"],  "env": {"PROVENZA_URL": "https://provenza.example.com", "PROVENZA_TOKEN": "pvz_enr_..."}}}}

On the first start the agent enrolls: its keys are made on that machine, and the token can then be removed. The client sees only the tools an administrator approved, and every call is checked against the organization's policies and recorded. Set PROVENZA_AGENT_NAME to run more than one agent on a machine.

Through the LLM gateway, the server checks allowed models and providers, policies and limits; it stores no prompt or answer. Every model the organization has routed (OpenAI, Anthropic, Azure, …) is reached through this one OpenAI-compatible endpoint. Streaming is not supported yet.

Tool wrappers. These keep the tool's name, description and argument schema, so the model sees no difference:

# pip install 'provenza[langchain]'from provenza.integrations.langchain import govern_toolstools = govern_tools(agent, [search, read_file])        # a refusal reaches the model as the tool's answer
# pip install 'provenza[autogen]'from provenza.integrations.autogen import govern_toolstools = govern_tools(agent, [FunctionTool(lookup, description="...")])
# pip install 'provenza[crewai]'from provenza.integrations.crewai import govern_toolsresearcher = crewai.Agent(role="...", goal="...", tools=govern_tools(agent, [search_tool]))

prefix= namespaces the names policies see, so prefix="kb." turns search into kb.search. For anything else, govern a single call with agent.call(name, arguments, lambda: ...) or await agent.acall(...).

What this is, and what it is not

The SDK runs inside your agent's process. It is the agent's honest record and cooperative enforcement: code in the same process could call a tool around it, and could read the agent's keys. Enforcement that does not depend on the agent's code comes from Provenza's gateways (LLM, MCP, A2A), which hold the credentials so the agent never has them.

Verify a local audit log with provenza.audit_log.verify(path, public_key). Verify a server receipt with tools/provenza_audit.py receipt.