Replio

An lightweight tooling core for fleets of single-purpose agents.

Replio is deliberately small and auditable zero-dependency agentic core built on a single streaming loop. The model plans, the tool registry acts, and the same loop powers an interactive REPL, headless CLI or HTTP API. Each process is a self-contained, scoped agent in one folder with its config, model and tool permissions. Light enough for one machine to hold a fleet of focused agents.

Features

• Zero dependencies - everything is Python standard library. Nothing to audit, no supply chain, no lockfile churn
• One agent loop - a single SSE stream per turn powers the REPL, the CLI and the API. No duplicated logic across front-ends
• Local-first - config and session logs live on your disk. Bring your own provider key, or run fully local
• Agentic REPL - streaming token-by-token output, dimmed thinking, markdown-aware rendering, readline history and tab completion
• Tool calling - web search, page fetching, file read/write/search and shell execution via OpenAI-compatible function calling
• Permissions - every tool is gated by allow / ask / deny, with path-scoped confirmations for anything outside your worktree
• Plugins - external repositories register tools, providers and slash commands. The core stays zero-dependency, and plugin deps are imported lazily, only when you activate a plugin
• Multi-provider - Ollama, OpenAI, Groq, Anthropic, plus any OpenAI-compatible endpoint, with automatic detection from the base URL
• Sessions - complete append-only conversation logs, including every tool call and its result
• Compaction - summarize long conversations and trim the provider context without losing history
• Headless modes - replio run for scripting and replio serve for an HTTP JSON API
• Agent fleets - one process per single-purpose agent, scoped to its own folder, config and permissions

Quick Start

pipx install replio
replio

Or from source:

git clone https://github.com/emyasnikov/replio.git && cd replio
python3 -m venv .venv && .venv/bin/pip install -e .
.venv/bin/replio

Usage

REPL

First-time setup with /connect, then type any message. Tab-complete / commands and session names. Use arrow keys to navigate history.

>>> /connect
  Provider [ollama]:
  Base URL [https://ollama.com]:
  API key: ...
  Model [gpt-oss:20b-cloud]:
>>> Hi
<<< Hello! How can I help you today?
>>> /exit

CLI

Stream plain text with --output text or return the results as JSON, log tool status and diagnostics to stderr with --verbose, and address a persistent session with --session-id <id>. Tools that require confirmation are auto-denied in by default, just pass --yes to approve them.

replio run --prompt "Hi"
{
  "content": "Hello! How can I help you today?",
  "thinking": null,
  "tool_calls": [],
  "errors": [],
  "duration": 7.0,
  "usage": null,
  "model": "gpt-oss:20b-cloud",
  "provider": "ollama",
  "session": "20260814_192251_hi",
  "status": "ok"
}

API

Server exposes JSON endpoints like POST /chat {"prompt": "..."} (optionally {"session_id": ...}) returns the same turn result as the CLI.

replio serve &
curl localhost:8787/chat -X POST -d '{"prompt": "Hi"}'
{"content": "Hello! How can I help you today?", "thinking": null, "tool_calls": [], "errors": [], "duration": 7.0, "usage": null, "model": "gpt-oss:20b-cloud", "provider": "ollama", "session": "20260814_192711_hi", "status": "ok"}

Agent orchestration

Three ways to combine Replio into larger systems:

• MCP (Model Context Protocol) - work alongside other AI tools. Connect to external MCP servers (stdio or HTTP) to import their tools, or expose Replio's own tools and sessions as an MCP server for other agents (Claude, opencode, ...) via replio mcp or POST /mcp on replio serve. See docs/mcp.md
• Swarm - make agents cooperate on a task with /agent personas and the delegate tool, so a lead agent splits work across specialized sub-agents and auditors. See docs/swarm.md
• Fleet - run many scoped agents side by side, each a replio serve process confined to its own folder, worktree and permissions, orchestrated under a supervisor. See docs/fleet.md

Fleet and swarm are two layers: fleet keeps agents alive, swarm gets the job done, and both compose with MCP for cross-tool interoperability.

Roadmap

Fleet orchestration (running many scoped agents under a supervisor) and swarm orchestration (agents cooperating through personas and delegation) are the two orchestration layers being built next. See docs/fleet.md and docs/swarm.md. Open tasks live in TODO.md.

Contributing

The project is stdlib-only with no external dependencies. See AGENTS.md for architecture and conventions, and CONTRIBUTING.md for the contribution workflow.

Documentation

Detailed references are in the docs/INDEX.md

License

MIT