This monorepo is the home of the OpenGuardrails (OGR) specification and its reference integrations. The specification is the normative contract every integration and detector speaks; the integrations, benchmark, examples, skill, and website live alongside it so changes can be reviewed and tested together.

OGR is not a guardrail product: it defines the wire and referees the leaderboard. Vendors compete on detection quality behind a common plug; users get one way to configure and compose safety & security across every agent they run.

• We define the wire — the layer model, events, verdicts, composition, taxonomy.
• We referee the benchmark.
• We do not build detection capability — vendors compete behind the contract.

The layer model: OGR beside OSI

This is the protocol's foundational concept. OGR is to agent traffic what the layered network model is to packets — and it is built the way a firewall is: an integration sees one event at a time, the way a firewall sees one IP packet, and the runtime reassembles everything above it and reads everything below it out of the payload.

Like a packet, an event is a header — kind (step/request | step/response), step_id, and the identity four-tuple agent_id · agent_type · agent_workspace · agent_user (OGR's answer to the firewall's 5-tuple) — plus a payload: the raw provider body. Everything above L3 is derived server-side (sessions by conversation-prefix chaining, turns by instruction boundaries and idle timeout — a firewall does not ask packets which connection they belong to); everything below is parsed from the payload (calls) or inferred (exec: no sensor observes it, and the gap between what a call claims and what an exec does is precisely what agent security is about).

Two honest notes on the analogy. OGR follows the pragmatic TCP/IP cut — a layer earns its place with its own unit, mechanism, and question — not OSI's seven: above transport, networking has only "application", but agent traffic is a dialogue with stable structure, so Turn and Session are this domain's own layers, defined here rather than mapped onto OSI's vestigial session/presentation layers. And the agent is an endpoint, not a layer — it persists with zero traffic, sessions belong to it the way TCP connections belong to a host, and it is addressed by the four-tuple every event carries. Beside the stack sits the entity axis every firewall has: tenant (the API key), workspace = security zone (one zone, one policy set), agent = host, discovered from traffic into an inventory.

Each event gets a verdict at the moment the integration can still refuse it — the request before the model sees it, the response before the agent acts on it:

  your own agent · harness plugins        gateway integrations
  (two POSTs at the loop's seams)         (an LLM proxy: Higress, …)
        │                                       │
        │   raw provider bodies + step_id       │
        ▼                                       ▼
   ┌───────────────────────────────────────────────┐
   │  OGR core contract                            │
   │  GuardEvent · Verdict ·                       │
   │  composition · taxonomy                       │
   └───────────────────────────────────────────────┘
                       ▲
                       │
                detector plugins
               (config rules OR model/classifier)

The same six layers, in five other vocabularies

Agent harnesses already have words for this traffic. They line up:

The numbers line up through L4 on purpose. Exec/call/event/step sit on physical/link/network/transport, and the packet is L3 in both columns. Above transport the columns part: networking has only "application", because network applications share no structure — agent traffic is a dialogue with stable structure, so turn and session are this domain's own L5 and L6, not OSI's session and presentation layers (the two practice discarded).

⚠️ "Turn" means this stack's STEP in two of the three SDKs. In both the OpenAI Agents SDK and the Claude Agent SDK a turn is one iteration of the agent loop — one model call plus the tool runs it triggers — and that is what max_turns counts. An OGR turn is the user-instruction episode that contains those iterations: one Runner.run(), one query() prompt, one graph invoke(). Same word, one layer apart. (The OpenAI Agents SDK documentation uses both senses: max_turns counts loop iterations, while "a single logical turn in a chat conversation" is one Runner.run() — an OGR turn.)

The full mapping — including what to send as session_hint, why an SDK hook (PreToolUse, wrap_tool_call) is an enforcement point where a tracing span is not, and how a handoff moves the entity axis rather than the stack — is in Overview § The layer model in harness vocabularies.

Normative text: Overview § The layer model.

Integrate your agent in five minutes

The whole protocol is one endpoint, two calls per model call. You forward the exact bodies you already send to and receive from your LLM; the runtime does everything else (sessions, turns, decomposition, detection). Fail-open by default: if the runtime is unreachable, your agent keeps running.

import uuid, requests

OGR = "https://ogr.example.com"           # your runtime's base URL
KEY = "ogr_xxxxxxxx"                      # your organization API key

# The identity four-tuple. All four always present; "" = nothing to assert
# (the runtime then derives identity from the API key).
IDENTITY = {
    "agent_id":        "invoice-bot",     # WHICH agent — unique in your org
    "agent_type":      "my-harness",      # what KIND — a label, never policy
    "agent_workspace": "finance-agents",  # agent GROUP — one policy set
    "agent_user":      "u-8232",          # who is USING it this session
}

SESSION = uuid.uuid4().hex   # optional session_hint: one id per conversation —
                             # sessions become declared instead of inferred

def evaluate(kind, step_id, payload):
    """The whole protocol is this one call. Fail-open: no verdict → proceed."""
    try:
        r = requests.post(f"{OGR}/v1/evaluate",
                          headers={"Authorization": f"Bearer {KEY}"},
                          json={"kind": kind, "step_id": step_id,
                                "llm_protocol": "openai.chat",
                                "session_hint": SESSION,
                                **IDENTITY, "payload": payload},
                          timeout=5)
        return r.json() if r.ok else None
    except requests.RequestException:
        return None

def blocked(v):
    return v is not None and v["decision"] == "block"

# your agent loop, with the two calls added:
while True:
    step_id = uuid.uuid4().hex                     # binds this call's 2 events
    body = {"model": "gpt-5", "messages": messages, "tools": TOOLS}
    if blocked(evaluate("step/request", step_id, body)):     # ① before the model
        break
    resp = call_llm(body)                                    # your code, unchanged
    if blocked(evaluate("step/response", step_id, resp)):    # ② before acting on it
        break
    ...                                            # execute tool calls, loop

Runnable version (with streaming): examples/minimal-agent/. Full contract: Runtime API — including a complete exchange: both halves of one model call written out whole, with the verdict each returns.

The questions the wire raises first — which llm_protocol to declare, what to send when your protocol is not one we list, why a different model does not mean a different integration — are answered in the protocols FAQ.

Why a standard

Without OGR, securing an agent is an N × M × L integration problem: every agent, every detector vendor, every LLM protocol wired pairwise. OGR collapses it to N + M + L — integrate once against the contract.

Two layers: API → Plugin

There is no SDK layer. The API is the integration surface — one decision endpoint and one recipe — and agent developers integrate by calling it directly:

The normative components

Risk categories live in the taxonomy (safety.* and security.*), versioned and swappable — the contract references category IDs but stays neutral on what is "unsafe."

Two domains, one contract

• Safety — harmful content/behavior (toxicity, self-harm, CSAM, brand, topic). Mostly classifier-judged at the content I/O boundary.
• Security — system compromise (prompt injection, data exfiltration, malicious commands, SSRF, secret leakage, supply chain). Judged on actions and data flow — what a tool call is about to do.

The contract is unified; the pipelines and enforcement points differ. Start with the overview.

Conformance & benchmark

• A detector is OGR-conformant if it accepts a GuardEvent and returns a valid Verdict against the JSON Schemas. See CONFORMANCE.md.
• The benchmark evaluates conformant detectors on shared corpora and publishes the leaderboard.

Monorepo layout

Integration status

The v0.6 SDK packages were retired in v0.7 — the API is the integration surface. v0.8 merged the two integration recipes into one, and every integration below speaks it (v1.0 releases the same wire unchanged):

Development

# benchmark tests
python -m pip install pytest && python -m pytest

# higress plugin
cd integrations/gateway/higress && go test ./...

# dsh plugin (npm workspace)
npm install && npm run build && npm test

Principles

1. Neutral. The protocol is open and foundation-governed; the benchmark is a referee, not a contestant.
2. Standardize the boundary, not the brains. Detection stays competitive.
3. Name the loop the way harnesses do. Session, turn, step, call — an integration should never have to translate its own vocabulary to speak the wire.
4. The wire carries what only the producer knows. Identity and the step-pairing id are asserted; everything derivable — sessions, turns, numbering, timestamps, protocol versions — is the runtime's job, so the integration stays stateless.

Status

Current protocol version: v1.10 (see CHANGELOG.md for protocol versions). The wire is the one v1.0 declared stable in 2026-08: everything since is additive-optional (additionalProperties: false rejects unknown keys, not absent ones, so both ends roll forward independently), so an integration written against v1.0 is conformant today and gains a key at a time. ⚠️ Each key ships at the RUNTIME first — a producer one version ahead sends a key a strict schema does not know, and the answer is a 400 on every event it sends. Anything breaking is a new major version. See GOVERNANCE.md for how the spec evolves. Contributions welcome — CONTRIBUTING.md.

License

Apache-2.0.