# Arcezia > Arcezia is AI agent security at the tool call: runtime security and action > authorization for AI agents. Before a tool call an agent makes runs (a database > write, a shell command, an email, a payment, an MCP tool call), Arcezia decides > whether it may run: ALLOW, REVIEW (held for a person, with what would release > it) or BLOCK (refused, with why). It decides from the call itself, the rules you > write and facts your own systems confirm, and keeps a signed record of every > decision that an auditor can check offline. It works alongside AI guardrails, > observability, evaluations, orchestration, identity and sandboxing, and is not > any of those. The same input always gets the same answer; under heavy load a call > may instead be answered 'not decided, try again', never a different verdict. One call checked three ways > on the live service, with the code to replay it: > https://arcezia.com/agent-action-authorization/ ## Who builds it - Founder: Praveen Babu, Chennai, India. Started in 2026; private beta; not yet incorporated; patent pending in India. https://arcezia.com/about/ ## What Arcezia does - Arcezia sits in front of an AI agent's tool calls (refunds, deployments, database writes, shell commands, API calls) and checks each one before it runs. Only ALLOW reaches the tool. - ALLOW lets the call run. BLOCK refuses it and says why. REVIEW holds it and names what would release it, such as a person's approval or an answer from your own check. - It checks whether an action is justified, not only whether the words are safe. A tool call can be polite, on-topic and well-formed and still rest on a fact no tool established. On our private suite (engine of the live build, 11 October 2026), of the 73 dangerous scenarios in which the agent supplied its own evidence, 71 were refused and 2 held for a person; none ran. - An approval the agent only claims is refused. A person's approval is a token your backend signs after the person clicks Approve. - Every answer comes with a signed record of the decision and its reason. It is evidence that the check ran before the action, not a certification. ## Definitions - Execution integrity: every action an AI agent executes is backed by evidence and authorized, checked before it runs rather than scored after. - Pre-execution verification: checking a tool call an AI agent proposes, and the facts it rests on, before the tool runs, then answering ALLOW, BLOCK or REVIEW. - Authority Fabrication Index (AFI): how often an AI agent asserts authority or evidence it does not actually have when attempting an action. - Full glossary: https://arcezia.com/glossary/ ## Pages - https://arcezia.com/ : what Arcezia does - https://arcezia.com/developer-docs/ : install the SDK, check a tool call, handle each answer; contracts, approvals, sessions, API, errors - https://arcezia.com/integrations/ : all framework hookups - https://arcezia.com/integrations/claude-code/ : a PreToolUse hook that checks Claude Code's shell commands, writes, web access and MCP tools - https://arcezia.com/integrations/langchain/ : LangChain and LangGraph tools that raise ToolException on REVIEW or BLOCK - https://arcezia.com/integrations/openai/ : OpenAI function calling - https://arcezia.com/integrations/anthropic/ : Anthropic tool use - https://arcezia.com/integrations/n8n/ : n8n workflows - https://arcezia.com/integrations/crewai-autogen-llamaindex/ : CrewAI, AutoGen, LlamaIndex and any Python function - https://arcezia.com/glossary/ : plain definitions: AI guardrails, agent runtime security, MCP security, agent governance, execution integrity, pre-execution verification, ALLOW/BLOCK/REVIEW, AFI, signed decision records - https://arcezia.com/agent-fabrication-benchmark/ : how often guardrails approve an action resting on invented evidence, and how to rerun the test - https://arcezia.com/fabricated-refund/ : a worked example, a refund for an order that never existed - https://arcezia.com/security/independent-review/ : an outside review of the live service, published in full and unchanged - https://arcezia.com/ai-agent-security/ : AI agent security in three parts (the model, the data, what the agent does), agent runtime security, AI security for agents, and how to control tool calls - https://arcezia.com/incidents/ : AI agent incidents from primary sources, with what the tool did and which layer failed - https://arcezia.com/incidents/2026-07-openai-agents-hugging-face/ : incident, 8 July 2026: OpenAI evaluation agents broke into Hugging Face production - https://arcezia.com/incidents/2026-04-pocketos-railway-volume-delete/ : incident, 24 April 2026: Cursor agent deleted PocketOS's production volume on Railway - https://arcezia.com/incidents/2026-02-openclaw-inbox-delete/ : incident, 23 February 2026: OpenClaw agent trashed an inbox it was told not to act on - https://arcezia.com/incidents/2026-02-lobstar-wilde-token-transfer/ : incident, 22 February 2026: Lobstar Wilde agent sent a thousand times more than intended - https://arcezia.com/incidents/2025-07-replit-production-database/ : incident, 18 July 2025: Replit agent deleted a production database in a code freeze - https://arcezia.com/incidents/2024-02-moffatt-air-canada-chatbot/ : incident, 14 February 2024: Air Canada chatbot misstated its bereavement fare policy - https://arcezia.com/banks/audit-procedure/ : audit procedure for AI agent action records at a bank: what to request, how to run the offline checker, how to sample - https://arcezia.com/banks/ : AI agent action records for banks: what auditors can check, mapped to the EU AI Act, DORA, PCI DSS, RBI directions and India's DPDP Rules - https://arcezia.com/healthcare/audit-procedure/ : audit procedure for AI agent action records in healthcare: what to request, how to run the offline checker, how to sample - https://arcezia.com/healthcare/ : AI agent action records for healthcare: what auditors can check, mapped to HIPAA, the EU AI Act, India's DPDP Act and Rules, and ABDM - https://arcezia.com/agent-action-authorization/ : AI agent action authorization (tool-call security, pre-execution control): one call checked three ways, ALLOW, REVIEW and BLOCK, with the SDK and HTTP code and the answers measured on the live service - https://arcezia.com/audit-evidence/ : AI agent audit evidence: a real signed sample export, the offline checker command and its measured output, and what PASS, PASS-ATTESTED, UNVERIFIED and FAIL mean - https://arcezia.com/where-arcezia-fits/ : where Arcezia fits next to guardrails, observability, evaluations, orchestration, workflow infrastructure, identity and sandboxing: what each layer answers, what it cannot answer before an action runs, and what Arcezia covers and does not - https://arcezia.com/ai-guardrails/ : AI guardrails and guardrails for AI agents: what text guardrails check, what they cannot see (the tool call and its effect), published measured comparison rows, and one runnable request that is refused - https://arcezia.com/mcp-security/ : MCP security: checking MCP tool calls in Claude Code before they reach the server, with the setup and two MCP calls answered on the live service - https://arcezia.com/agent-governance/ : AI agent governance per action: a rule, a decision before each tool call, a signed record and an offline check, run end to end on the live service ## Setup (for coding assistants) Full recipes with expected output: https://arcezia.com/developer-docs/#setup 1. Install: `pip install -U arcezia cryptography requests` (Python 3.10 or newer). 2. Key: an owner key from the Keys page at https://app.arcezia.com/. `export ARCEZIA_API_KEY=""`. For a key starting `ar_test_`, also `export ARCEZIA_ENV=production`. 3. Signing key: make an Ed25519 key pair and keep the private half where the agent cannot read it. Register the public half: `az.register_token_key(public_key)` or `POST /v1/account/token_key {"public_key": ""}`. Expect `ok: true` (then `GET /v1/account/token_key` shows `registered: true`). 4. Contract: one per tool, for example `{"tools": {"execute_sql": {"pack": "database_ops", "not_present": ["outbound", "trust_boundary_crossing", "sensitive_data"]}}}` via `az.register_contract(contract, name="setup")` or `POST /v1/contracts {"name", "contract"}`. Expect `domains["execute_sql"]`: always check that tool under this domain. 5. Session: a signed envelope, for example `{"allowed_domains": [], "allowed_action_types": ["execute_sql"], "resource_scope": ["orders", "42"]}`, via `arcezia.Arcezia(task=..., signing_key=private_key).start_session(capability_envelope=...)` or `POST /v1/session {"task", "capability_envelope_token"}`. Expect a `session_id`. 6. Check each call before it runs: `az.verify(action_type, action_description, domain)` or `POST /v1/verify {"task", "session_id", "action_type", "domain", "action_description"}`. Expect `verdict`: ALLOW (run it), REVIEW (`release` names what would release it), BLOCK (`reason` names why). 7. First answers to expect: `SELECT status FROM orders WHERE id = 42` -> ALLOW; `SELECT * FROM orders` -> REVIEW ['approval:user']; the same tool under `database_ops` -> BLOCK ['contract:tool_pairing']. 8. Claude Code: `pip install -U arcezia`, `arcezia-hook install`, then write `~/.claude/arcezia.json` as `{"capability_envelope": {"allowed_domains": [...], "allowed_action_types": [...]}}`. In `~/.claude/settings.json`, add `"onFailure": "block"` to the Arcezia PreToolUse entry (Claude Code 2.1.295 or later), so a missing or crashed hook blocks instead of letting calls through. The hook answers allow, ask (REVIEW) or deny (BLOCK). 9. A REVIEW naming `check:` (for example a write: REVIEW ['approval:production', 'approval:user', 'check:verified_recent_backup']): register an HTTPS endpoint of yours under exactly that name, `POST /v1/probes {"domain", "constraint_name", "webhook_url", "secret"}` (201, returns `signing_key_v2` once). Any other name is refused with 400 `invalid_probe`. A check that times out, errors or cannot tell keeps the call held (REVIEW `check:`); a check's no can refuse it (BLOCK `check:`); a check for a person's approval can only refuse, never approve. ## Public repositories - https://github.com/Praveendante/arcezia-python : the Python SDK source (MIT) - https://github.com/Praveendante/arcezia-bench : benchmark cases, harness and method you can rerun against the hosted service - https://github.com/Praveendante/agent-incident-ledger : a public record of AI agent incidents, one file per incident, primary sources only (CC-BY 4.0), maintained by Arcezia's founder, Praveen Babu (Chennai, India) ## Notes for machine readers - The SDK is `arcezia` on PyPI (Python 3.10 or newer). Any language can call the HTTP API. The checking service is hosted, not distributed. - Integrations: Claude Code hook, LangChain, LangGraph, OpenAI function calling, Anthropic tool use, CrewAI, AutoGen, LlamaIndex, n8n. - Status: private beta. SOC 2: not yet started.