CategoryWeightCategory scoreContribution

Securityprompt, exec, net, exfil, eval

35%

26.3 pts

Supply chainhash, typosquat, maintainer, lockfile

20%

100

20.0 pts

Maintenancestaleness, pinning, CI

15%

100

15.0 pts

TransparencySKILL.md, perms, README

15%

100

15.0 pts

Communityinstalls, verify, response

15%

100

15.0 pts

Findings & checks · 1 flagged

Securityscore 75 · 1 finding

HIGHLong base64-encoded blob hidden in the skill documentationSS-SKILL-INJECT-B64-PAYLOAD-01 · Prompt injection · tests/e2e/protocol/scenarios/claude-code/image-basic.json

HIGH — once decoded by the agent, an encoded payload has the same impact class as plain-text injection.

Why it matters

A base64 string of 128+ characters appears in a documentation file. Encoded prompt injection hides the hostile instruction in base64 — invisible to keyword filters — and relies on the agent's ability to decode it at runtime. There is no normal authoring reason to embed a multi-hundred-byte base64 blob in skill docs.

The exact value spotted

excerpttests/e2e/protocol/scenarios/claude-code/image-basic.json· json

29"type": "base64",

30"media_type": "image/png",

31"data": "iVBORw0KGgoAAAANSUhEUgAAAAIAAAACCAYAAABytg0kAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQU

… (86 chars elided on L31)

32}

33}

Occurrences

1 occurrence · at L31

How to fix

Remove the encoded blob, or decode it and review what it actually contains.

Decode the base64 string and confirm it is not an instruction directed at the agent.
Move any legitimate binary or signature data into a dedicated file (*.sig, SIGNATURES) outside the documentation.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-B64-PAYLOAD-01sha256a4c2b7c147052f5b…rubric 365aacaView on GitHub

Supply chainscore 100 · 0 findings

✓—All supply chain checks passedNo findings in this category for the latest scan.pass

Maintenancescore 100 · 0 findings

✓—All maintenance checks passedNo findings in this category for the latest scan.pass

Transparencyscore 100 · 0 findings

✓—All transparency checks passedNo findings in this category for the latest scan.pass

Communityscore 100 · 0 findings

✓—All community checks passedNo findings in this category for the latest scan.pass

The primary manifest — the file an agent reads to learn what this artifact does.

README.md · 1 file

README.md8.3 KB · Markdown

CliGate

CliGate Dashboard

[English](./README.md) | [中文](./README_CN.md)

CliGate is a local AI control plane built around two core capabilities:

Assistant: a resident private assistant that stays available in the background, understands user tasks, remembers context, uses tools, schedules work, operates channels, and can execute actions through MCP, skills, desktop automation, shell/file tools, or delegated runtimes.
Model Proxy: one local API and model-routing layer for Claude Code, Codex CLI, Gemini CLI, OpenClaw, and API-compatible clients, with unified accounts, API keys, model mapping, logs, usage, and cost visibility.

It keeps both layers local-first on localhost: the assistant acts like a personal operator for real tasks, while the proxy owns provider access, routing, credentials, unified model names, and observability.

Why CliGate

One local dashboard for a resident private assistant and unified model routing
Assistant tasks with memory, approvals, follow-ups, scheduled work, channels, and tool execution
Account pools, API keys, local runtimes, app routing, and model mapping in one proxy layer
Channels for Telegram, Feishu, and DingTalk workflows
Local-first deployment without a hosted relay

What It Includes

Assistant

Dashboard chat and Assistant Tasks for personal task execution
A persistent assistant agent with task records, memory, policies, approvals, and resumable executions
Tool execution through skills, MCP, shell/file tools, scheduled tasks, channels, and optional desktop automation
Optional delegation to Codex / Claude Code runtime sessions when a task needs an external coding agent
Telegram, Feishu, and DingTalk channel workflows

Model Proxy

Anthropic Messages, OpenAI Chat Completions, OpenAI Responses, Codex, and Gemini-compatible endpoints
One-click configuration for Claude Code, Codex CLI, Gemini CLI, and OpenClaw
ChatGPT, Claude, and Antigravity account pools
API key pools for OpenAI, Azure OpenAI, Anthropic, Gemini, Vertex AI, MiniMax, Moonshot, ZhipuAI, DeepSeek, Qwen, and OpenRouter
Routing priority, app-level bindings, provider model mapping, and free-model routing
Optional local model routing through Ollama-style runtimes

Observability and operations

Usage and pricing views
Request logs and live log streaming
API explorer
Tool installer and CLI config helpers
Resources catalog for free/trial model providers

Quick Start

1. Start CliGate

npx cligate@latest start

Or install globally:

npm install -g cligate
cligate start

Or use a desktop release package:

Download the installer or app package for your platform from Releases.
Install or open the package, then launch CliGate.
CliGate will start the local service and open the desktop window automatically.

Default dashboard:

http://localhost:8081

2. Add at least one working credential

Use the dashboard:

Accounts for ChatGPT / Claude / Antigravity
API Keys for provider keys
Local Models for on-device runtimes

3. Choose your first path

For Assistant use, open Chat or Assistant Tasks and tell the private assistant what you want done.

For Model Proxy use, point a CLI tool or API-compatible client to CliGate.

Claude Code:

export ANTHROPIC_BASE_URL=http://localhost:8081
export ANTHROPIC_API_KEY=any-key
claude

Codex CLI:

# ~/.codex/config.toml
chatgpt_base_url = "http://localhost:8081/backend-api/"
openai_base_url = "http://localhost:8081"

Gemini CLI and OpenClaw can also be configured from the dashboard.

User Paths

Assistant users

Use Chat, Assistant Tasks, Conversation Records, Scheduled, Skills, MCP, and channels to ask the resident assistant to execute real tasks, remember context, use tools, send follow-ups, and keep working in the background.

Model Proxy users

Start the service, add one credential, run one-click config, and send your first proxied request from Claude Code, Codex CLI, Gemini CLI, OpenClaw, or an API-compatible client.

Dashboard operators

Use the dashboard to manage accounts, API keys, routing priority, model mapping, local runtimes, pricing, request logs, usage, channel settings, skills, MCP, and desktop-agent settings.

Screenshots

Dashboard	Chat
Dashboard	Chat

Routing and Settings	Channels
Settings	Channels

Usage and Costs
Usage and Costs

Documentation

Start here if you want the shortest path to the right document:

After the server starts, a lightweight product guide is also available at:

http://localhost:8081/manual/
http://localhost:8081/resources/

Local Architecture

Assistant Surfaces
  Web Chat / Assistant Tasks / Telegram / Feishu / DingTalk / Scheduled Tasks
           |
           v
Private Assistant and Tools
  Memory / Policies / Skills / MCP / Desktop Agent / Shell + File Tools / Optional Codex + Claude Code Delegation
           |
           v
CliGate Local Control Plane (localhost:8081)
           |
           +--> Model Proxy
           |    - Protocol translation
           |    - Account and API key routing
           |    - App-level bindings and model mapping
           |    - Local model routing
           |
           v
Upstream Providers and Local Runtimes
  OpenAI / Anthropic / Gemini / Vertex AI / Kilo / Ollama / others

API Surface

Endpoint	Use
`POST /v1/messages`	Anthropic Messages proxy
`POST /v1/chat/completions`	OpenAI Chat Completions proxy
`POST /v1/responses`	OpenAI Responses proxy
`POST /backend-api/codex/responses`	Codex internal compatibility
`POST /v1beta/models/*`	Gemini CLI proxy
`GET /api/agent-runtimes/providers`	Runtime provider catalog
`GET /api/agent-channels/conversations`	Channel conversation records
`GET /api/assistant/tasks`	Assistant task records
`GET /api/assistant/mcp/servers`	MCP server management
`GET /api/assistant/skills`	Assistant skill management
`GET /api/desktop-agent/status`	Desktop-agent status
`GET /api/local-runtimes`	Local runtime status
`GET /api/resources`	Resource catalog
`GET /health`	Health and version

See docs/API.md for more detail.

Community

For Chinese-speaking users, you can also add the maintainer directly on personal WeChat:

Please include a short note such as CliGate when sending the request.

If you plan to contribute, read CONTRIBUTING.md before opening a pull request.

License

This project is licensed under AGPL-3.0.

Disclaimer

CliGate is an independent open-source project and is not affiliated with Anthropic, OpenAI, Google, or other upstream providers.

cligate— agent skill

Is cligate safe to install?

CliGate

Why CliGate

What It Includes

Assistant

Model Proxy

Observability and operations

Quick Start

1. Start CliGate

2. Add at least one working credential

3. Choose your first path

User Paths

Assistant users

Model Proxy users

Dashboard operators

Screenshots

Documentation

Local Architecture

API Surface

Community

License

Disclaimer

Audit the pieces. Scan the whole. Decide.