slife 0.1.3

Terminal-based AI agent — a function-calling loop with minimum harness

slife

Terminal-based AI agent — a function-calling loop with minimum harness. Chat with an LLM that can execute shell commands, search the web, load on-demand skills, connect to MCP servers, call any REST API via OpenAPI specs, spawn subagents for parallel work, and communicate with other slife instances over MQTT.

Quick Start

uv sync
cp slife.json5.example slife.json5
# Edit slife.json5 — set your LLM provider's API key
uv run slife

The example config includes three pre-configured MCP servers (filesystem, fetch, duckduckgo-search) that need no API keys — you're ready to go after setting your model key.

Configuration

Edit slife.json5. See slife.json5.example for the full annotated template — it covers models, env vars, MCP servers, and the commented github REST API template. The minimum you need:

• Provider + API key — set models.providers.<name>.api_key via ${ENV_VAR} or inline
• Active model — active_model: "provider/model-id"
• MCP servers — pre-configured with filesystem, fetch, and duckduckgo-search (no auth needed)

Tools

All tools are unified as OpenAI function definitions — the LLM sees no difference between them.

Native Functions

Auto-discovered from slife/tools/. Use slife.json5 only to override defaults or disable a tool.

Tool	What it does
execute_shell	Execute a shell command with configurable timeout
run_python_script	Platform-correct Python invocation with JSON args
get_os_info	Return current OS: Windows, Linux, or macOS
config_env_set / get / remove	Manage env vars in slife.json5 + os.environ
cli_add_tool / check_installed / remove / list	Register, check, and discover external CLIs

Skills

On-demand documentation plugins under skills/. The agent loads them only when needed via list_skills / use_skill. Each skill is a directory with a SKILL.md file. Install new skills at runtime with add_skill / remove_skill.

MCP & REST APIs

External MCP servers connect through slife-mcp — an independent proxy that manages persistent connections. Tools are prefixed by server name (e.g. filesystem__read_file). REST APIs connect the same way via anyapi-mcp-server, which converts any OpenAPI spec to callable tools at runtime.

Add servers at runtime with mcp_add_server or pre-configure them in slife.json5 → mcp.servers. Servers default to eager mode (all tools loaded at startup). For servers with many tools, use disclosure: "lazy" — the server connects but tools load on demand via mcp_set_disclosure, keeping context lean.

A2A — Agent-to-Agent

Two transports, one interface. The full A2A protocol toolset (14 tools) provides discovery, task routing, lifecycle management, and notifications:

Tool	Role
a2a_list_agents	List all agents on the MQTT mesh (includes self)
a2a_list_subagents	List local subagent workers
a2a_send_task	Send a task and wait for the result (sync)
a2a_send_task_async	Fire-and-forget, returns task ID for polling
a2a_get_task_result	Poll task status and result from TaskStore
a2a_list_tasks	List all tasks with status/agent/transport filters
a2a_cancel_task	Cancel a pending or in-flight task
a2a_subscribe_task	Block until a task completes (event-driven or poll)
a2a_agent_card	Introspect a specific agent's status
a2a_spawn_subagent	Create a local worker with the same LLM + tools
a2a_stop_subagent	Stop a locally-managed subagent
a2a_notify_user	Fire a desktop notification to the human operator
a2a_broadcast	Scatter/gather — send a task to all known agents

Transport	Enable	Use case
MQTT	--name <id> CLI flag	Remote slife instances (P2P mesh)
Subagent (stdin/stdout)	Always available	Local child processes for parallel work

Start with --name my-agent to join the MQTT mesh; subagents are always available.

When --name is provided the agent identity flows through the entire UI:

• Prompt prefix changes from > to my-agent> for your own messages.
• System prompt includes your name so the LLM knows its identity.
• Remote tasks from other agents appear as other-agent> task… and stream responses to the chat view just like locally-typed messages.
• Log files are named logs/slife_my-agent_YYYYMMDD_HHMMSS.log for easy identification in multi-agent sessions.

Tips

• /file image.png — attach an image for vision models
• /exit — quit the application
• Ctrl+L — clear the conversation
• Esc — focus the input field

Design

slife is a minimum-harness agent. The harness only does what the LLM cannot: execute tools, maintain conversation state, and stream responses. The system prompt contains only project-specific facts not in the LLM's training data. See DESIGN.md for the full rationale and architecture.

Project Structure

slife/
  agent/            # LLM client, conversation, function-calling loop, inbox
  a2a/              # A2A: MQTT client, broker lifecycle, identity, TaskStore
  subagent/         # Subagent: spawn, JSON-RPC IPC, process management
  tools/            # All tools — native, skills, CLI, A2A (auto-discovered)
    a2a.py          #   14 A2A protocol tools (unified transports)
  mcp/              # MCP client (slife side)
  ui/               # Textual TUI
slife_mcp/          # Independent MCP proxy (pip install slife-mcp)
skills/             # On-demand skill plugins
tests/              # pytest suite (583 tests)

Requirements

• Python ≥ 3.13
• uv (Python package manager)
• Node.js (only if using npx-based MCP servers)

License

MIT