AgentKit

中文版本

AgentKit is a general-purpose framework for tool-using LLM agents. It provides workspace isolation, a unified LLM abstraction, a tool-execution loop, and both CLI and Python SDK entry points.

Current Capabilities

• Strong workspace isolation: all agent-visible file access goes through WorkspaceFS
• Unified LLM abstraction: ConversationItem / ConversationState / UnifiedLLMRequest / UnifiedLLMResponse
• Provider support (sync, non-streaming, text)
  • OpenAI (responses / chat_completions)
  • Anthropic (messages)
  • Gemini (generateContent)
  • Qwen (OpenAI-compatible chat_completions)
  • vLLM (OpenAI-compatible chat_completions)
• Registry-based tool system with built-in tools under src/agentkit/tools/library/ and custom tool loading from project file/directory entries
• Agent loop that coordinates model reasoning and tool execution under step/time budgets
• Run logs written to workspace/logs/run_<run_id>.jsonl

Status

• The repository includes unit tests for the agent loop, CLI, tool system, runlog, and the OpenAI / Anthropic / Gemini / Qwen / vLLM provider adapters
• There is no smoke-test coverage against live upstream APIs yet, so a final end-to-end validation against the target model is still recommended before production use

Documentation

Full documentation lives under docs/agentkit/. The published site is GitHub Pages and deploys automatically from pushes to main.

Install From Source

uv sync
uv run agentkit --help

Install From PyPI

pip install base-agentkit
agentkit --help

The import path and CLI command stay the same after installation:

from agentkit import create_agent

Example Configuration

vLLM

workspace:
  root: "./vllm_workspace"

provider:
  kind: "vllm"
  model: "glm-5"
  openai_api_variant: "chat_completions"
  conversation_mode: "auto"
  base_url: "http://localhost:8000/v1"
  api_key: "empty"
  temperature: 0.8
  timeout_s: 600
  retries: 2
  enable_thinking: true

agent:
  system_prompt: "You are a helpful agent. Use tools when needed."
  budget:
    max_steps: 200
    time_budget_s: 1800
    max_input_chars: 180000

tools:
  entries:
    - "./tools"
  allowed:
    - "view"
    - "create_file"
    - "str_replace"
    - "word_count"

runlog:
  enabled: true
  redact: true
  max_text_chars: 20000

Run From The CLI

export OPENAI_API_KEY="your-key"
uv run agentkit --config path/to/config.yaml run --task "List the files in the current workspace"

Python SDK

from agentkit import create_agent

agent = create_agent("path/to/config.yaml")
report = agent.run("Create notes/todo.txt in the workspace and write today's plan")
print(report.final_output)

Project Layout

The core implementation lives under src/agentkit/:

• config/: configuration schema and loader
• workspace/: strict filesystem isolation and workspace layout
• llm/: unified abstractions, provider base/factory, and adapters
• tools/: tool abstractions, registry, and autoloading
• agent/: budgets, reports, runtime, and the agent loop
• runlog/: structured events, event sinks, and JSONL runlogs
• cli/: command-line entry point