Sage Agent

Sage Agent

Yes, I shamelessly named it after me ;)

Inspired by the recent sprawl of OpenClaw, PicoBot, ZeroClaw, and whatever else popped up last Tuesday — I decided to write my own. Written from the ground up in Python.

Sage doesn't aspire to be the next Claude Code. Instead, it's intentionally designed to be a clean slate out of the box, so that you can make it more intelligent. No opinions. No bloat. Just a solid foundation you can build on top of.

Built-in evaluation and CI/headless execution are included. See [.docs/eval.md](.docs/eval.md) and [.docs/ci-headless.md](.docs/ci-headless.md).

Key Features

Agents

The core unit. Define an agent in a Markdown file with YAML frontmatter — name, model, system prompt — and you're running. No boilerplate classes, no framework ceremony. Just config and go.

---
name: assistant
model: gpt-4o
---
You are a helpful AI assistant.

Subagents & Delegation

Agents can have subagents. When they do, they automatically get a delegate tool — the LLM decides when and how to hand off work. It's orchestration without the orchestration code.

The delegate tool accepts two optional parameters beyond the target agent and task:

• session_id — resume a previous conversation with the subagent. Conversation history is persisted across calls and restored automatically. The result is prefixed with [Session: ] when a session ID is supplied.
• category — route the delegation to a model defined in the categories block of config.toml (e.g. "quick" → gpt-4o-mini, "deep" → Claude Sonnet). The subagent's model is swapped at runtime without changing its config file.

Tools via @tool Decorator

Write a Python function. Decorate it with @tool. Sage auto-generates the JSON schema from your type hints. That's it. No manual schema wrangling.

@tool
def word_count(text: str) -> str:
    """Count the number of words in the given text."""
    return str(len(text.split()))

Built-in tools included — or load them all at once with sage.tools.builtins:

| Category | Tools | |----------|-------| | Core | shell, file_read, file_write, file_edit, http_request | | Memory | memory_store, memory_recall | | Web | web_fetch, web_search | | Git | git_status, git_diff, git_log, git_commit, git_undo, git_branch |

Per-agent tool restrictions are supported via frontmatter. blocked_tools hides specific tools from the LLM; allowed_tools is an explicit allowlist (all others are hidden). Blocklist takes precedence over allowlist.

---
blocked_tools: [shell, http_request]
allowed_tools: [file_read, file_write, memory_store]
---

Skills

Reusable capabilities defined as Markdown files. Drop them in a skills/ directory and all agents share them automatically. Sage resolves the global skill pool via a waterfall (skills_dir in config.toml → ./skills/ → ~/.agents/skills/ → ~/.claude/skills/). Each agent can optionally limit its skills to a named subset via an allowlist in config.toml. Flat files or directory-per-skill — both work.

Orchestration

Four flavors:

• Pipeline (>>) — chain agents sequentially. Output of one feeds the next.
• Parallel — run multiple agents concurrently via Orchestrator.run_parallel().
• Race — first agent to complete wins via Orchestrator.run_race().
• Autonomous delegation — an orchestrator agent with subagents decides who does what, on its own.

100+ LLM Providers

Powered by litellm. OpenAI, Azure, Anthropic, Ollama, and basically everything else. One model string, any provider.

| Provider | Model String | |----------|-------------| | OpenAI | gpt-4o, gpt-4o-mini | | Azure | azure/gpt-4o | | Anthropic | anthropic/claude-sonnet-4-20250514 | | Ollama | ollama/llama3 |

MCP Support

Connect to MCP servers (stdio or SSE) or expose your tools as an MCP server. Both directions work.

Semantic Memory

SQLite-backed with litellm embeddings. Zero-config persistent recall across sessions. Compaction built in so context doesn't bloat forever.

Permissions

Control what tools can do via a single permission: block in YAML frontmatter. Each permission category (read, edit, shell, web, memory) maps to a set of built-in tools. Set a category to allow, deny, or ask, or use pattern matching for fine-grained shell control. When set to deny, tools are invisible to the LLM. Interactive prompts in the TUI when policy is ask.

Hook System

A lifecycle event bus for intercepting and extending agent behavior without modifying core code. Register async handlers against named HookEvent values — 31 events across seven categories: run lifecycle (ON_RUN_STARTED, ON_RUN_COMPLETED, ON_RUN_FAILED, ON_RUN_CANCELLED), LLM calls (PRE_LLM_CALL, POST_LLM_CALL, ON_LLM_STREAM_DELTA, ON_LLM_ERROR, ON_LLM_RETRY), tools (PRE_TOOL_EXECUTE, POST_TOOL_EXECUTE, ON_TOOL_FAILED), delegation (ON_DELEGATION, ON_DELEGATION_COMPLETE, ON_DELEGATION_FAILED), memory, compaction, permissions, sessions, coordination, and planning. Built-in hooks cover credential scrubbing, query-based model routing, bail-out retry (follow-through), automatic memory injection, notepad injection, and plan analysis. Hooks that raise never crash the agent — errors are logged and swallowed.

from sage.hooks.registry import HookRegistry
from sage.hooks.base import HookEvent

hr = HookRegistry()

async def log_calls(event, data):
    print(f"{event}: {data.get('model')}")

hr.register(HookEvent.PRE_LLM_CALL, log_calls)
agent = Agent(name="a", model="gpt-4o", hook_registry=hr)

Every hook emission is also recorded as a canonical EventEnvelope by the telemetry layer (sage/telemetry.py). Each event carries correlation IDs (run_id, session_id, originating_session_id), timing, token usage, and a sanitized payload. The TUI's event timeline and inspector consume these envelopes via JSON-RPC for real-time visibility into agent behavior. See [ADR-012](.docs/adrs/012-event-telemetry-and-observability.md).

Coordination

Agent-to-agent messaging and lifecycle primitives for multi-agent systems:

• MessageBus — in-memory per-agent inboxes with TTL expiry, idempotency, overflow protection, and broadcast delivery
• CancellationScope — propagate cancel signals across async tasks; child scopes inherit parent cancellation
• SessionManager — create, track, and destroy concurrent agent sessions with typed metadata
• BackgroundTaskManager — launch subagent runs as non-blocking asyncio tasks. The orchestrator gets a delegate_background tool and receives completion notifications injected into the next turn's message stream. Supports polling (background_status) and cancellation (background_cancel).

# Orchestrator gets these tools automatically when subagents are present:
# delegate_background(agent_name, task) → task_id
# background_status(task_id) → status + result
# background_cancel(task_id) → bool

Context Management

Token-aware context window management. Automatic compaction when approaching the model's limit — tries LLM summarization first, then emergency drop, then deterministic trim as a guaranteed last resort. Configurable reserve tokens and optional pruning of large tool outputs.

TUI

A full interactive terminal UI built with Ink v6 and React 19. Communicates with the Python backend via JSON-RPC over stdio. Block-based conversation display with live streaming, collapsible tool calls, markdown rendering, permission prompts, delegation hierarchy visualization, and an event timeline with inspector for real-time observability.

Prerequisites: Node.js 22+, pnpm 10+

# Install dependencies
make tui-install

# Build
make tui-build

# Install globally on PATH
make tui-install-global

# Development mode (hot reload)
make tui-dev

See [tui/README.md](tui/README.md) for slash commands, keyboard shortcuts, and architecture details.

Planning Pipeline

A structured plan-then-execute loop for long-horizon tasks. Enable it via planning: in frontmatter.

• plan_create — agents call this tool to create a named plan with an ordered list of task descriptions. State persists to disk under .sage/plans/.
• plan_status / plan_update / plan_complete — query and mutate plan state across turns and sessions.
• notepad_write / notepad_read — persistent markdown working memory scoped to a plan. Notes are stored under .sage/notepads// and injected automatically before each LLM call via the built-in notepad hook.
• Plan analysis — optional ON_PLAN_CREATED hook that makes an LLM call to identify ambiguities, missing dependencies, ordering issues, and risks in a newly created plan. Enable via planning.analysis.enabled: true.
• Review loop — review_loop(plan, reviewer, reviser) iterates through review-revise cycles using the PlanReviewer protocol. The shipped LLMPlanReviewer evaluates specificity, success criteria, and dependencies. Configurable via planning.review.
• ConductorMixin — mixin for orchestrator agents that drives plan execution: reads the plan, delegates each pending task to an executor subagent, and persists results after each step.

---
name: planner
model: gpt-4o
planning:
  analysis:
    enabled: true
  review:
    enabled: true
    max_iterations: 3
---
You are a planning agent. Use plan_create to structure work, then execute step by step.

Prompt System

Model-specific overlays — lightweight transformations applied to the assembled system prompt after all content (body, identity, skills) has been joined. Each overlay targets a model family and appends model-tuned instructions.

Built-in overlays:

• GeminiOverlay — appends a tool-call enforcement reminder for gemini/* models
• GPTOverlay — appends a "format reasoning in clear steps" hint for gpt-* models

Register custom overlays via overlay_registry.register(my_overlay) from sage.prompts.

Dynamic delegation table — orchestrator agents can include {{DELEGATION_TABLE}} in their system prompt body. Sage replaces this placeholder at runtime with a markdown table of all available subagents, derived from each agent's prompt_metadata frontmatter field.

---
name: orchestrator
model: gpt-4o
subagents: [researcher, summarizer]
---
You are the orchestrator.

{{DELEGATION_TABLE}}

Use the delegate tool to assign work.

# researcher/AGENTS.md
---
name: researcher
model: gpt-4o
description: "Finds primary sources and synthesizes citations"
prompt_metadata:
  cost: moderate
  use_when: ["deep research", "fact checking", "citations needed"]
  avoid_when: ["simple questions", "quick lookups"]
  triggers: ["research", "find sources", "citations"]
---

Protocol-Based Architecture

ProviderProtocol, MemoryProtocol, EmbeddingProtocol — swap out any layer. Don't like the SQLite memory backend? Write your own. Don't want litellm? Implement the protocol. Everything is async-first.

Quick Start

pip install sage-agent
# or
uv tool install sage-agent

export OPENAI_API_KEY=sk-...
sage agent run AGENTS.md --input "What is the capital of France?"

Code API

import asyncio
from sage import Agent

agent = Agent(
    name="assistant",
    model="gpt-4o",
    body="You are a helpful assistant.",
)

result = asyncio.run(agent.run("What is 2 + 2?"))
print(result)

Or load from config:

agent = Agent.from_config("AGENTS.md")
result = asyncio.run(agent.run("Hello"))

Pipelines

pipeline = researcher >> summarizer
result = asyncio.run(pipeline.run("Explain quantum computing"))

Parallel Execution

from sage import Orchestrator

results = asyncio.run(Orchestrator.run_parallel(agents, "Analyze this topic"))

Race Execution

winner = asyncio.run(Orchestrator.run_race(agents, "Solve this problem"))

Autonomous Orchestration

---
name: orchestrator
model: gpt-4o
subagents:
  - research_agent
  - summarize_agent
---
You are an orchestrator. Use the delegate tool to assign tasks to your subagents.

sage agent run orchestrator/AGENTS.md --input "Research and summarize quantum computing"

CLI

sage agent run AGENTS.md --input "Hello" [--stream]   # Run an agent
sage agent validate AGENTS.md                          # Validate config
sage agent list [directory]                            # List agent configs
sage agent orchestrate AGENTS.md --input "text"        # Run subagents in parallel
sage tool list AGENTS.md                               # List available tools
sage init [--name my-agent] [--model gpt-4o]           # Scaffold a new project
sage tui --agent-config AGENTS.md                      # Launch interactive TUI
sage exec AGENTS.md -i "Hello" [-o text|jsonl|quiet] [--timeout N] [--yes]    # Run headless (CI/scripting)
sage eval run suite.yaml [--min-pass-rate 0.9] [--runs N]                      # Run evaluation suite
sage eval validate suite.yaml                                                   # Validate suite file
sage eval history [--suite NAME] [--last N]                                     # Show run history
sage eval compare                                           # Compare two runs
sage eval list [directory]                                                       # Find suite files

Configuration Reference

Agent Config (Markdown Frontmatter)

---
name: my-agent
model: gpt-4o
description: "A helpful assistant"   # Display only, NOT sent to model
max_turns: 10
max_depth: 3                       # Max delegation depth (default: 3)

git:
  auto_snapshot: true              # Auto-snapshot before edits (default: true)

# Tool access: permission categories drive tool registration
# Categories: read, edit, shell, web, memory, task, git
# Values: "allow" | "deny" | "ask" | {pattern: action, ...}
permission:
  read: allow
  edit: allow
  shell:
    "*": ask
    "git log*": allow
    "git diff*": allow
  web: allow

# Custom tool modules (in addition to permission-derived built-ins)
extensions:
  - myapp.tools                       # Your own tools (module path)

memory:
  backend: sqlite                    # "sqlite" (default) or "file"
  path: memory.db
  embedding: text-embedding-3-large
  compaction_threshold: 50
  auto_load: false                   # Auto-inject recalled memories pre-LLM-call
  auto_load_top_k: 5                 # How many memories to inject

subagents:
  - research_agent                   # Directory containing AGENTS.md
  - config: helper.md                # Reference another .md file
  - name: inline-helper              # Or define inline
    model: gpt-4o-mini

mcp_servers:
  filesystem:
    transport: stdio
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
  remote:
    transport: sse
    url: http://localhost:8080/sse

context:
  compaction_threshold: 0.75         # Compact at 75% of context window
  reserve_tokens: 4096
  prune_tool_outputs: true
  tool_output_max_chars: 5000

model_params:
  temperature: 0.7
  max_tokens: 2048

# Hook-driven features (all optional)
credential_scrubbing:
  enabled: true
  patterns: ["sk-.*", "Bearer .*"]
  allowlist: ["sk-test"]

query_classification:
  rules:
    - pattern: "python|code"
      model: gpt-4o
      priority: 1

follow_through:
  enabled: true
  patterns: ["I cannot", "I'm unable", "I don't have access"]

research:
  enabled: true
  max_sources: 3
  timeout: 10.0

session:
  enabled: true

# Per-agent tool restrictions
blocked_tools: [shell, http_request]   # always hidden from the LLM
# allowed_tools: [file_read, memory_store]  # allowlist (all others hidden)

# Planning pipeline (Phase 3)
planning:

…

## Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [sagebynature](https://github.com/sagebynature)
- **Source:** [sagebynature/sage-agent](https://github.com/sagebynature/sage-agent)
- **License:** MIT
- **Homepage:** https://pypi.org/project/sage-agent

Install and usage instructions live in the source repository linked above.