chimera-runtime

Deterministic Runtime for AI Agents
Stop trusting LLMs. Start constraining them.

Quickstart • Constraints • Integrations • Dashboard • How It Works

---

The Problem

LLMs don't follow rules. They approximate them.

They hallucinate. They get prompt-injected. They break constraints silently. And yet, we hand them real-world decisions -- financial approvals, infrastructure changes, customer-facing actions -- and hope for the best.

Hope is not a runtime strategy.

The Solution

Chimera is a deterministic execution layer that sits between your AI agent and the real world.

Not a wrapper. Not a filter. Not another prompt engineering trick.

A runtime that enforces constraints outside the model, verifies policies with formal logic (Z3), and blocks invalid actions before they happen.

Every decision is: Checked. Proven. Enforced.

Agent Intent → Constraint Guard → ALLOW / BLOCK / ASK_HUMAN → Audit Record

No exceptions. No bypass. No "the model said it was fine."

---

Quickstart

pip install chimera-runtime

chimera-runtime init
chimera-runtime run

Three commands. Your agent is now constrained.

---

Define Constraints

Two paths. Same enforcement.

Option A: YAML Rules (zero dependencies)

# policies/governance.yaml
rules:
  - name: manager_limit
    when: "role == 'MANAGER' and amount > 250000"
    then: BLOCK
    message: "Managers cannot approve more than $250,000"
  - name: weekend_freeze
    when: "is_weekend == 'YES' and urgency != 'CRITICAL'"
    then: BLOCK
    message: "No changes on weekends unless critical"

Option B: CSL with Z3 Formal Verification (recommended)

CONFIG {
  ENFORCEMENT_MODE: BLOCK
}

DOMAIN GovernanceGuard {
  VARIABLES {
    amount: 0..1000000
    role: {"ANALYST", "MANAGER", "DIRECTOR", "VP", "CEO"}
  }

  STATE_CONSTRAINT manager_approval_limit {
    WHEN role == "MANAGER"
    THEN amount <= 250000
  }
}

YAML gives you rules. CSL gives you mathematical proof that your constraints are consistent, reachable, and conflict-free -- before a single agent runs.

chimera-runtime verify policies/governance.csl

---

Framework Integrations

Chimera drops into whatever you're already using. No rewrites.

LangChain

from langchain_openai import ChatOpenAI
from chimera_runtime.integrations.langchain import wrap_tools, ChimeraCallbackHandler

guarded_tools = wrap_tools(your_tools, policy="./policies/governance.yaml")

handler = ChimeraCallbackHandler(policy="./policies/governance.yaml")
llm = ChatOpenAI(callbacks=[handler])

LangGraph

from chimera_runtime.integrations.langgraph import compliance_node

graph.add_node("constraint_gate", compliance_node(policy="./policies/governance.yaml"))

CrewAI

from chimera_runtime.integrations.crewai import wrap_crew_tools

guarded_tools = wrap_crew_tools(your_tools, policy="./policies/governance.yaml")

Direct Usage

from chimera_runtime import ChimeraAgent

agent = ChimeraAgent(
    model="gpt-4o",
    api_key="sk-...",
    policy="./policies/governance.csl",
)

result = agent.decide(
    "Approve $200k marketing spend for Q3",
    context={"role": "MANAGER", "department": "MARKETING"},
)

print(result.result)       # "ALLOWED"
print(result.action)       # "Allocate $200k to digital channels"

Every tool call. Every decision. Through the constraint guard. No exceptions.

---

Dashboard

Full runtime control plane at compliance.chimera-protocol.com.

Real-time decision enforcement -- Live feed of every ALLOW, BLOCK, and ESCALATE. See exactly what your agents are doing and what they're being stopped from doing.

Policy management with formal verification -- Write, edit, and verify CSL policies. Z3 proves correctness before deployment. No untested constraints reach production.

Decision analytics & block rate heatmaps -- What's getting blocked, when, and why. Patterns across agents, frameworks, and time.

Framework integration wizard -- Four steps to wire Chimera into LangChain, LangGraph, CrewAI, LlamaIndex, or AutoGen.

---

How It Works

1. Define constraints in CSL
2. Z3 verifies correctness (reachability, consistency, conflict-freedom)
3. Runtime enforces at execution time

The model never sees the constraints. The model never evaluates the constraints. The model's opinion on whether it should be allowed to do something is irrelevant.

Constraints live outside the model. Enforcement happens outside the model. The model is a generator. Chimera is the governor.

---

CLI Reference

Command	Description
chimera-runtime init	Initialize project with config and starter policy
chimera-runtime run	Interactive agent with real-time constraint enforcement
chimera-runtime run --daemon	Pipe mode for JSON stdin/stdout
chimera-runtime stop [--force]	Graceful or immediate halt
chimera-runtime verify [POLICY]	Verify CSL policy (syntax + Z3 + IR)
chimera-runtime test [--skip-llm]	End-to-end system test
chimera-runtime audit --stats	Aggregate decision statistics
chimera-runtime audit --last 10	Recent decisions table
chimera-runtime audit --export FILE	Export records (json/compact/stats)
chimera-runtime policy new NAME	Create policy from template
chimera-runtime policy list	List and verify all policies
chimera-runtime policy simulate FILE '{"k":"v"}'	Test policy against input
chimera-runtime explain --id ID	Generate Art. 86 HTML explanation
chimera-runtime docs generate	Generate technical documentation
chimera-runtime license status	Show current license tier
chimera-runtime license activate KEY	Activate a license
chimera-runtime license deactivate	Remove license key

---

EU AI Act

Compliance is a consequence -- not the product.

When you enforce deterministic constraints, maintain complete audit trails, and provide human oversight by design, you don't need to bolt on compliance after the fact. You already have it.

Article	Requirement	How Chimera Delivers It
Art. 9	Risk management	Formal verification (Z3) proves policy correctness before deployment
Art. 12	Record-keeping	Complete DecisionAuditRecord for every decision, automatically
Art. 13	Transparency	Full decision explanations, audit queries, HTML reports
Art. 14	Human oversight	ASK_HUMAN escalation, confirm/override/halt with audit trail
Art. 15	Accuracy & security	Deterministic constraint gate -- not probabilistic filtering
Art. 19	Retention	Configurable retention with automatic enforcement
Art. 86	Right to explanation	Self-contained HTML reports per decision
Annex IV	Technical documentation	Auto-generated (14/19 sections)

---

Supported Frameworks & Providers

Integration	Install	Notes
LangChain	pip install chimera-runtime[langchain]	Tool wrapper + callback handler
LangGraph	pip install chimera-runtime[langgraph]	Node guard + state checkpoint
LlamaIndex	pip install chimera-runtime[llamaindex]	Tool spec + callback handler
CrewAI	pip install chimera-runtime[crewai]	Tool wrapper for crews
AutoGen	pip install chimera-runtime[autogen]	Agent wrapper
OpenAI	pip install chimera-runtime[openai]	gpt-4o, gpt-4o-mini, o1, o3
Anthropic	pip install chimera-runtime[anthropic]	claude-sonnet-4-20250514, claude-opus-4-20250514
Google	pip install chimera-runtime[google]	gemini-2.0-flash, gemini-2.5-pro
Ollama	pip install chimera-runtime[ollama]	llama3, mistral, qwen (local)
CSL-Core	pip install chimera-runtime[csl]	Z3 formal verification engine

---

Development

git clone https://github.com/akarlaraytu/chimera-runtime
cd chimera-runtime
pip install -e ".[dev,all]"
pytest tests/ -v

---

License

Apache License 2.0 -- see LICENSE for details.

---

LLMs generate. Chimera governs.
_{Built by Aytug Akarlar}