Ailice
A self-hosted general-purpose assistant that decomposes complex requests into collaborating agent tasks.
- Source repo
- myshell-ai/AIlice
- Stars
- ★ 1.4k
- Last updated
- 1y ago
- License
- MIT
- Primary language
- Python
- FA score
- 50/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformOpenAI API · Claude API
- You'll need
- Typical use
- An individual who wants to run research, document reading, and result synthesis locally or in a container.
- Main limitation
- The default Google module is rate-limited; frequent use requires Custom Search JSON API credentials and google-api-python-client.
- Source review
- 50/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
Ailice is an autonomous general-purpose assistant built around open-source LLMs and served through a browser-based conversation interface. Its IACT, or Interactive Agents Call Tree, model lets AProcessor instances create sub-agents, request missing information from callers, and return results through the call tree. LLM output is interpreted by core/Interpreter.py into function calls, variable operations, and other actions, while independent processes provide modules such as search, browsing, code execution, and vector storage. Configuration in config.json supports either one model for all agents or separate models for distinct agent types. It can run locally or in Docker, and can expose stdio or SSE MCP servers as extension modules.
Users start the service with ailice or ailice_web and submit work through the web dialogue interface. The main agent can delegate to sub-agents; during multi-round execution, AProcessor calls peripheral modules and Interpreter parses matching LLM text into actions. The embedded syntax supports variable definition, variable references, text concatenation, and function calls; detected code blocks can be stored as variables and passed to an execution module. The services section in config.json can launch standalone module processes, while ailice_mcp_wrapper connects local stdio MCP servers or SSE MCP servers. Results are returned through the active conversation and agent-call chain.
- An individual who wants to run research, document reading, and result synthesis locally or in a container.
- A developer who needs an assistant to carry out multi-step programming, script execution, or system-management tasks.
- A team that wants to assign different models to the main, search, and other agent types through config.json.
- An integration developer who needs to connect a self-hosted stdio or SSE MCP server to an assistant workflow.
- A local deployment user who needs voice dialogue, rich media handling, and file upload/download in the web interface.
How do you install or deploy this agent?
Prepare Python and Chrome. Run:
git clone https://github.com/myshell-ai/AIlice.git
cd AIlicepip install -e .
For optional PDF reading, speech, Hugging Face model, or fine-tuning features, install only the needed extras: pip install -e .[pdf-reading], pip install -e .[speech], pip install -e .[huggingface], or pip install -e .[finetuning]. For an isolated deployment, run docker build -t ailice . followed by docker run -it -p 127.0.0.1:5000:5000 --name ailice ailice --expose=1 --contextWindowRatio=0.2.
How do you use this agent?
On first run, Ailice asks for an API key; it can also be changed in the config.json location printed at startup. Start with:
ailece --contextWindowRatio=0.2The documented executable is ailice, so use:
ailece is not a documented command; use ailice --contextWindowRatio=0.2.
Open the supplied web page and enter a task. A documented model-specific invocation is ailice --modelID=anthropic:claude-sonnet-4-20250514 --contextWindowRatio=0.2. If modelID is omitted, Ailice uses per-agent assignments under agentModelConfig in config.json. Run ailice --help for parameters. To add MCP tools, configure a services entry using ailice_mcp_wrapper, such as a stdio wrapper for mcp_echo hello.
What are this agent's strengths and limitations?
- Its IACT call tree allows a sub-agent to ask its caller for clarification or help, rather than relying on one-way function calls.
- External capabilities are independent processes that can run in different software, hardware, or remote environments.
- The repository documents MCP integration for both stdio and SSE servers through ailice_mcp_wrapper.
- It supports commercial providers, OpenRouter/Apipie, local Hugging Face models, and OpenAI-compatible inference services.
- Per-agent model selection makes it possible to combine models according to capability, cost, and local hardware limits.
- The default Google module is rate-limited; frequent use requires Custom Search JSON API credentials and google-api-python-client.
- For complex local tasks, the documentation recommends models of 70B parameters or larger and says effective use needs at least two RTX 4090 GPUs with 48GB VRAM.
- Ubuntu has the strongest support; Windows has limited testing and no integrated Windows command-execution tools, with Docker or WSL recommended.
- Initial use of open-source models, speech recognition, or TTS may download weights and require substantial time and disk space.
- Public share links are temporarily removed for security, and the documentation warns against exposing the app.py service to untrusted networks.
How does this agent compare with similar options?
Ailice can use official OpenAI, Mistral, Anthropic, and Groq model paths directly, or route models through OpenRouter and Apipie. For local or non-OpenAI-compatible inference, the documentation provides configuration paths for LM Studio and for Ollama converted with LiteLLM.
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| Ailice This agent | 50 · Major gaps | ★ 1.4k | 1y ago | Python | OpenAI API · Claude API |
| Keinsaas Navigator | 69 · Some gaps | ★ 1.2k | 1mo ago | TypeScript | OpenAI API · Claude API |
| Conductor Workflow Engine | 68 · Some gaps | ★ 32k | 1d ago | Java | Claude Code · OpenAI API · Claude API |
| Hermes Agent | 67 · Some gaps | ★ 248k | today | Python | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: The project offers sandbox run options (Docker) and recommends running in isolated environments; users can configure API keys, but no fine-grained permission control or user confirmation mechanism is provided. Data flow transparency is limited, with no detailed explanation of how data is processed or stored. Sensitive data handling is unclear, with no explicit statement on user data or API key handling. Dependency security is not fully addressed, with a broad dependency list but no vulnerability scanning or version pinning. External effects include code execution, software installation, and network access, but no rollback mechanism is provided. Source attribution is clear from the author, but the publisher is unverified. Deductions: lack of explicit least-privilege implementation, insufficient user confirmation, opaque data flow, unclear sensitive data handling, insufficient dependency security, no rollback for external effects, and attribution based only on author claims.
Evidence shows: The project description is consistent with functionality, and the architecture design is detailed, but no automated test results are provided. Dependency availability is not fully guaranteed, with a broad dependency list but no version pinning or mirrors. Failure messages are insufficient, with no detailed error handling or user guidance. Deductions: good self-consistency, but dependency availability is not fully ensured and failure messages are lacking.
Evidence shows: The project targets various user scenarios with rich examples and use cases. Capability boundaries are clear, supporting multiple LLMs and module extensions. Trigger precision relies on natural language instructions, but no precise trigger mechanism is provided. Environment fit supports multiple OS and Docker, but Windows support is limited. Deductions: clear capability boundaries, but trigger precision depends on LLM understanding and environment fit has limitations.
Evidence shows: Information architecture is clear, README is detailed, and installation instructions are complete. Naming is stable, and version numbers are explicit. Examples and FAQ are rich, and known limitations are partially mentioned. License is MIT, but versioning changelog is missing. Maintenance responsibility is clear from the author, but no contribution guide is provided. Deductions: good information architecture, but missing changelog and insufficient maintenance responsibility.
Evidence shows: Output usability is high, with rich UI and multiple output formats. Marginal value is high, providing autonomous agent capabilities. Cost-benefit is affected by hardware requirements for local LLM running. Deductions: good output usability and marginal value, but cost-benefit is impacted by hardware requirements.
Evidence shows: Claims in README are partially supported by examples, but no independent verification is provided. Cross-source corroboration is limited, with no third-party evaluation. Fact-inference separation is not explicitly distinguished. Deductions: insufficient claim traceability, limited cross-source corroboration, and unclear fact-inference separation.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- The project allows arbitrary code execution and software installation, posing security risks; recommend running in isolated environments.
- Handling of sensitive information such as API keys is unclear; configure with caution.
- The dependency list is broad without version pinning, posing supply chain risks.
- Publisher identity is unverified; evaluate with caution.