Feather

Feather is a personal AI agent that lives in your terminal. You talk to it, it talks back, and it can use tools (read your files, run shell commands, search the web, remember things across sessions) to help you get work done.

The package is published on PyPI as feather-agent-os. The command you type is just feather.

What it does

• Chat with a single "lead" agent through a colorful terminal UI (built on Textual) or a simpler streaming console.
• Use built-in tools for reading files, running bash, searching code with grep, fetching web pages, and more.
• Optionally turn on long-term memory so the agent remembers facts about you across sessions.
• Spawn helper sub-agents in the background for bigger jobs (research, validation, exploration).
• Add your own skills, custom sub-agents, and MCP servers without touching the package code.
• Reach the agent from Telegram, LINE, or WhatsApp.
• When running on OpenRouter, opt in to broadcast every turn's metadata (session, agent, model) to Comet Opik or any other observability destination OpenRouter supports. See docs/providers.md.

How it fits together

%%{init: {"flowchart": {"htmlLabels": true, "padding": 16, "nodeSpacing": 60, "rankSpacing": 70}, "themeVariables": {"fontSize": "16px"}}}%%
flowchart TB
    USER["You"]

    subgraph FRONT["Front doors"]
        TUI["Textual TUI / Rich CLI"]
        WH["Webhook Router<br/>Telegram, LINE, WhatsApp"]
    end

    USER --> TUI
    USER --> WH

    LEAD["Lead Agent<br/>BaseAgent.run_loop"]

    TUI --> LEAD
    WH --> LEAD

    subgraph CAPS["Capabilities"]
        TOOLS["Built-in tools<br/>read_file, grep, bash,<br/>web_search, ask_user, ..."]
        SUBS["Sub-agents<br/>explore, research, validate,<br/>your custom agents"]
        SKILLS["Skill catalog<br/>loaded on demand"]
        CRON["Scheduler<br/>cron and once jobs"]
        MEM["Long-term memory<br/>Qdrant + Gemini<br/>optional"]
        MCP["MCP servers<br/>stdio and http<br/>optional"]
    end

    LEAD --> TOOLS
    LEAD --> SUBS
    LEAD --> SKILLS
    LEAD --> CRON
    LEAD --> MEM
    LEAD --> MCP

    PROV["Provider<br/>OpenAI Responses or OpenRouter Chat"]
    LEAD --> PROV
    SUBS --> PROV

    subgraph STORE["Storage"]
        SQL["SQLite<br/>sessions, tasks, cron,<br/>agent mailbox"]
        FILES["File overflow<br/>.feather/tmp and attachments"]
        QD["Qdrant<br/>vector store"]
    end

    LEAD --> SQL
    SUBS --> SQL
    CRON --> SQL
    MEM --> QD
    TOOLS --> FILES

The full architecture, with sequence diagrams for the agent loop, sub-agent dispatch, memory pipeline, skills, MCP, and compaction, is in docs/architecture.md.

60-second quickstart

pip install feather-agent-os                    # or: pipx install feather-agent-os
                                          # or: uv tool install feather-agent-os
feather init-memory                       # optional: start Qdrant so the
                                          # agent remembers things across
                                          # chats (needs Docker)
feather onboard                           # one-time setup, collects keys
feather                                   # opens the chat UI

Type a question. Press Enter to send. Type /exit when you are done.

If you do not have an OpenAI key yet, get one at https://platform.openai.com/api-keys. If you ran feather init-memory, you will also need a Gemini key from https://aistudio.google.com/apikey for embeddings; the wizard will ask for it. The full memory walkthrough is in docs/memory.md.

Skip feather init-memory if you do not have Docker. The agent still works fine, it just starts every chat fresh.

Where to go next

Where Feather stores things

Two locations, on purpose. Personal stuff (API keys, your profile, installed skills) lives in your home directory and follows you across projects. Project stuff (chat history for one repo, attachments, artifacts) lives next to your code.

Run feather init inside any folder to pin a project to that folder. Otherwise Feather walks up from where you are and uses the first .feather/ it finds; if it finds none, it falls back to the global location.

The full layout and override env vars are documented in docs/configuration.md and docs/getting-started.md.

Develop locally

git clone https://github.com/timho102003/feather-os.git
cd feather-os
uv sync                                   # editable install plus dev deps
uv run feather --version
uv run pytest

The version is derived from git tags via hatch-vcs. uv build produces the wheel and sdist under dist/.

License

Apache License 2.0. See LICENSE.

The full changelog is in CHANGELOG.md.