AI Agents You CanActually Trust
The open-source AI agent protocol built for auditability, safety, and control.
Drop-in alternative to OpenClaw and MCP. Every tool call is scored, every action is logged, every credential is encrypted.
Why ClearFrame?
Existing AI agent frameworks have critical security and auditability gaps
OpenClaw / MCP Problems
- • Single process reads untrusted content AND executes tools → prompt injection
- • Credentials stored in plaintext
~/.env - • No audit trail — forensics impossible
- • No concept of what the agent is supposed to do
- • Chain-of-thought never captured
- • No operator control plane
ClearFrame Solutions
- • Reader/Actor isolation — two sandboxed processes
- • Encrypted Vault — AES-256-GCM, memory-locked
- • HMAC-chained Audit Log — tamper-evident
- • Goal Monitor — alignment scoring, auto-pause on drift
- • Reasoning Transparency Layer — full trace as JSON
- • AgentOps — live REST + WebSocket dashboard
Core Features
Production-grade security and observability for AI agents
Reader/Actor Isolation
Two sandboxed processes communicate over a typed pipe. The reader never executes tools; the actor never reads untrusted input.
Encrypted Vault
AES-256-GCM encrypted credential storage. Memory-locked, auto-zeroizes on session end. No more plaintext API keys.
Tamper-Evident Audit Log
HMAC-chained audit trail. Cryptographically verify nothing was tampered with using clearframe audit-verify.
Goal Monitor
Every tool call is scored for alignment with the declared goal. Auto-pause on drift detection.
Reasoning Transparency
Full chain-of-thought captured as hash-verified JSON traces. Replay with clearframe rtl-replay.
AgentOps Dashboard
Live REST + WebSocket control plane. Approve, block, or tweak agent actions in real-time at localhost:7477.
Quick Start
Get started in minutes with a simple, declarative API
import asyncio
from clearframe import AgentSession, ClearFrameConfig
from clearframe.core.manifest import GoalManifest, ToolPermission
async def main():
config = ClearFrameConfig()
manifest = GoalManifest(
goal="Search for the latest AI safety papers and summarise them",
permitted_tools=[
ToolPermission(tool_name="web_search", max_calls_per_session=5),
],
)
async with AgentSession(config, manifest) as session:
result = await session.call_tool("web_search", query="AI safety 2026")
print(result)
asyncio.run(main())pip install clearframepython agent.pyclearframe startHow Does It Compare?
ClearFrame vs OpenClaw vs MCP
| Feature | OpenClaw | MCP | ClearFrame |
|---|---|---|---|
| Reader/Actor isolation | |||
| Goal alignment scoring | |||
| Reasoning trace capture | Partial | Full JSON | |
| Tamper-evident audit log | |||
| Encrypted credential vault | |||
| Context feed hashing | |||
| Live operator control plane | |||
| Signed plugin registry | |||
| Auto-pause on drift | |||
| Open source |
Ready to Build Trustworthy AI Agents?
Install ClearFrame and start building with confidence
pip install clearframe