🕵️‍♂️ Overview

AgentP (O.W.C.A. Top Secret Privacy Shield) is a high-performance, local reverse proxy that protects sensitive secrets, API keys, and PII from ever leaving your machine when interacting with cloud LLMs.

Modern AI coding agents (such as OpenCode, Claude Code, Cursor, and Antigravity) frequently inspect your workspace, .env files, and terminal logs. If prompt context contains database passwords, private keys, or API tokens, they get broadcast to upstream model providers.

AgentP solves this with Bi-Directional Fake-Swapping:

1. Outgoing Prompt Interception: Before requests leave your computer, AgentP scans prompts using an on-device HuggingFace NER model combined with deterministic regex rules.
2. Syntactic Fake Insertion: Real secrets are replaced with realistic dummy tokens (sk_live_fake..., sk-ant-fake...) preserving exact key syntax and prefixes.
3. Consistent State Cache: If an API key appears multiple times across turns or roles (system, user, assistant, tool), the exact same fake key is reused.
4. Incoming Response Restoration: When the upstream LLM returns code or tool calls containing the dummy tokens, AgentP swaps the real keys back in memory before handing the response to your agent.

⚡ Quick Start

Option 1: Run instantly with npx (Zero Installation)

npx @tamgallc/agentp

Option 2: Global CLI Installation

npm install -g @tamgallc/agentp
agentp

Option 3: Python Package

git clone https://github.com/Uunan/agentp.git
cd agentp
pip install -r requirements.txt
pip install torch --index-url https://download.pytorch.org/whl/cu126  # Optional: for GPU acceleration

# Launch setup wizard & proxy:
agentp

🤖 1-Click Coding Agent Setup

AgentP includes automated configuration installers for all major AI coding agents across Linux, macOS, and Windows:

# Configure all agents in a single step:
agentp --install all

# Or configure specific agents:
agentp --install opencode     # OpenCode (~/.config/opencode/opencode.json)
agentp --install claude       # Claude Code (~/.claude/settings.json)
agentp --install cursor       # Cursor (User/settings.json)
agentp --install antigravity  # Antigravity (~/.gemini/antigravity-cli/settings.json)

Supported Coding Agent Details:

| Agent | Target Config File | Configuration Applied | | :--- | :--- | :--- | | OpenCode | ~/.config/opencode/opencode.json | Configures agentp provider using @ai-sdk/openai-compatible pointing to http://127.0.0.1:8000/v1 | | Claude Code | ~/.claude/settings.json | Sets env.ANTHROPIC_BASE_URL to http://127.0.0.1:8000 (uses /v1/messages Anthropic adapter) | | Cursor | AppData/Roaming/Cursor/User/settings.json | Configures cursor.openai.baseUrl to http://127.0.0.1:8000/v1 | | Antigravity | ~/.gemini/antigravity-cli/settings.json | Sets openai_base_url to http://127.0.0.1:8000/v1 and injects AGENTP_PROXY_URL |

🛠️ How It Works (Architecture)

sequenceDiagram
    autonumber
    actor Dev as Developer / IDE
    participant Agent as Coding Agent<br/>(OpenCode / Cursor / Claude)
    participant AgentP as 🕵️‍♂️ AgentP Proxy<br/>(Localhost:8000)
    participant Upstream as Cloud LLM<br/>(NVIDIA / OpenAI / Anthropic)

    Dev->>Agent: Prompt contains Stripe/AWS/OpenAI secret key
    Agent->>AgentP: POST /v1/chat/completions (Real secrets)
    Note over AgentP: 1. Sliding window NER + Regex scan<br/>2. Real secret replaced with sk_live_fakeXYZ<br/>3. Mapping saved in ephemeral RAM
    AgentP->>Upstream: POST /v1/chat/completions (Fake secrets only)
    Note over Upstream: LLM sees ONLY fake tokens.<br/>Zero confidential data exposed.
    Upstream-->>AgentP: Response containing sk_live_fakeXYZ
    Note over AgentP: 4. Recursive JSON & tool-calls swap-back<br/>5. sk_live_fakeXYZ restored to Real Key
    AgentP-->>Agent: Final code containing real working key
    Agent-->>Dev: Execution succeeds flawlessly

🔌 API Endpoints & Compatibility

AgentP acts as a drop-in proxy replacement for both OpenAI and Anthropic API specifications:

| Method | Endpoint | Description | | :--- | :--- | :--- | | GET | /health | Service health status and active NER model check | | GET | /v1/models | OpenAI-compatible model discovery endpoint (for OpenCode and Cursor) | | POST | /v1/chat/completions | Standard OpenAI chat completions endpoint (supports SSE streaming and tool calling) | | POST | /v1/messages | Native Anthropic Messages API adapter (for Claude Code) | | POST | /v1/responses | Responses-API proxy endpoint | | POST | /detect | Standalone PII/Secret detection endpoint for verification and debugging |

💻 CLI Commands & Options

agentp                   # Starts proxy (opens interactive setup wizard on first run)
agentp --setup           # Launches the interactive configuration wizard
agentp --install all     # Configures OpenCode, Claude Code, Cursor, and Antigravity
agentp --port 8080       # Starts proxy on custom port
agentp --host 0.0.0.0    # Binds to custom host interface
agentp --check           # Runs local NER model smoke test

🔒 Security & Privacy Guarantees

• 100% On-Device Detection: The HuggingFace token-classification model runs entirely on your local CPU or GPU. Prompts are never transmitted to third parties for PII analysis.
• Strict Ephemeral RAM Mapping: Real-to-fake translation dictionaries live strictly in memory and are discarded immediately when the request lifecycle ends. No sensitive values are written to disk, SQLite, or log files.
• Sliding-Window Coverage: Prompts and codeblocks of arbitrary length (including documents far exceeding 256 tokens) are scanned without truncation.
• Deterministic Regex Fallback: In addition to deep NER, specialized patterns catch naked API keys (Stripe, OpenAI, Anthropic, AWS, GitHub PATs, Slack, NVIDIA Cloud) with 100% deterministic precision.

📄 License & Credits

Built by Tamga LLC under the Apache-2.0 License.
Special agent operations managed by O.W.C.A. Division.