inzen-bot 0.3.2

Multi-provider AI chatbot for the terminal

inzen-bot

A multi-provider AI chatbot for the terminal. Chat with Claude, GPT-4, Ollama, or llama.cpp without leaving your shell — with persistent conversation history, named sessions, token cost tracking, and a full interactive REPL mode.

No browser. No switching apps. Just ask and keep working.

---

Contents

• Features
• Install
• Setup
• Usage
• Interactive Mode
• Named Sessions
• Token Usage
• Providers
• Configuration
• Security
• Requirements
• Contributing
• Releasing to PyPI

---

Features

• Multi-provider — Claude, GPT-4, Ollama, llama.cpp, Groq, and any OpenAI-compatible endpoint
• Persistent history — follow-up with bot step 2? and it knows exactly what you mean
• Named sessions — keep separate threads for different projects, all stored locally
• Interactive REPL — bot --chat for a persistent conversation without re-typing flags
• Token & cost tracking — per-message footer and a bot --usage summary table across all providers
• Session export — dump any session to markdown or plain text
• Secure by default — local data files are user-only (600/700), session names are validated against path traversal

---

Install

pip install inzen-bot

Or with uv:

uv add inzen-bot

---

Setup

1. Set your API key

inzen-bot reads your API keys from environment variables.

Linux / macOS — add to your ~/.bashrc or ~/.zshrc:

export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."        # optional

Then reload:

source ~/.bashrc

Windows (PowerShell) — set permanently:

[System.Environment]::SetEnvironmentVariable("ANTHROPIC_API_KEY", "sk-ant-...", "User")

Restart your terminal after running this.

For Ollama and llama.cpp no API key is needed — just make sure the local server is running before use.

2. Verify it works

bot hello

You should see a streamed response from Claude with a token usage footer. If you get an error, check your API key:

# Linux / macOS
echo $ANTHROPIC_API_KEY

# Windows PowerShell
echo $env:ANTHROPIC_API_KEY

---

Usage

Ask anything — history is automatic so you can follow up naturally:

bot how do I format a drive in Ubuntu?
bot step 2?
bot step 3?

Common flags:

# Use a specific provider for one query
bot --provider openai explain RAID levels

# Switch default provider permanently
bot --set-provider ollama

# Switch model
bot --set-model claude-opus-4-5

# View conversation history
bot --history

# Purge data older than the configured retention window
bot --purge

# Clear history
bot --clear

# List all configured providers
bot --providers

# Override the system prompt for one query
bot --system "You are a Python expert" explain decorators

# Redact likely secrets in history/export output
bot --history --redact-secrets

# Export outside safe directories (disabled by default for safety)
bot --export myproject --output /tmp/myproject.md --unsafe-output

---

Interactive Mode

Use --chat for a persistent REPL — no need to retype flags between messages.

# Start a persistent chat session
bot --chat

# Start with an opening message, then continue the conversation
bot --chat explain how systemd services work

# Use a named session in interactive mode
bot --session myproject --chat

Inside the chat session the following slash commands are available:

Command	Description
/help	Show available commands
/history	Print the conversation so far
/clear	Clear conversation history for this session
/exit	Exit chat mode (also: /quit, q)

---

Named Sessions

Named sessions let you maintain separate, persistent conversation threads — useful for keeping different projects or topics isolated.

# Start or continue a named session
bot --session myproject what is a venv

# Pick up where you left off
bot --session myproject how do I activate it

# List all saved sessions
bot --sessions

# Export a session to markdown
bot --export myproject

# Export to a file
bot --export myproject --output myproject.md

# Delete a session
bot --clear-session myproject

Sessions are stored as JSON files in ~/.bot/sessions/ and are independent of provider — you can switch providers mid-session.

---

Token Usage

Every response includes a per-message token and cost footer. Run bot --usage to see cumulative totals across all providers:

# View cumulative usage across all providers
bot --usage

---

Providers

Provider	Requires	Notes
anthropic	ANTHROPIC_API_KEY	Default. Claude models.
openai	OPENAI_API_KEY	GPT models. Also works for Groq, Together, etc.
ollama	Ollama running	Local inference. Free to run.
llamacpp	llama.cpp running	Local inference via OpenAI-compatible server.

Ollama quickstart

# Install from https://ollama.ai, then:
ollama pull llama3
ollama serve
bot --set-provider ollama
bot hello from the terminal

llama.cpp quickstart

# Build llama.cpp from https://github.com/ggerganov/llama.cpp, then run the server:
./server -m your-model.gguf --port 8080
bot --set-provider llamacpp
bot what is quantisation

Using Groq or other OpenAI-compatible APIs

Point the OpenAI provider at any compatible endpoint via ~/.bot/config.json:

{
  "providers": {
    "groq": {
      "model": "llama-3.3-70b-versatile",
      "api_key_env": "GROQ_API_KEY",
      "base_url": "https://api.groq.com/openai/v1"
    }
  }
}

---

Configuration

Config and history live in ~/.bot/, with user-only file permissions on Linux and macOS:

~/.bot/                           (chmod 700)
├── config.json                   (chmod 600) — provider settings and defaults
├── history_anthropic.json        (chmod 600) — conversation history per provider
├── history_openai.json           (chmod 600)
├── history_ollama.json           (chmod 600)
├── usage_anthropic.json          (chmod 600) — cumulative token/cost data
└── sessions/                     (chmod 700)
    └── myproject.json            (chmod 600)

History is capped at 50 message pairs per provider. You can edit ~/.bot/config.json directly to customise models, add providers, or set custom base URLs.

Security-related defaults in config.json:

{
  "security": {
    "safe_output": true,
    "redact_secrets": false,
    "warn_dangerous_commands": true,
    "allowed_export_dirs": ["."],
    "allowed_hosts": [],
    "allow_insecure_http": false,
    "request_timeout_seconds": 30,
    "max_retries": 2,
    "retention_days": 30
  }
}

---

Security

• File permissions — ~/.bot/ and all data files are created with user-only permissions (700/600) on Linux and macOS so other users on the same machine cannot read your history, API keys, or session data.
• Session name validation — session names are validated against a strict allowlist (letters, numbers, ., _, -) to prevent path traversal attacks. Names like ../etc/passwd are rejected at the CLI boundary.
• Safe exports — --output writes are restricted to allowed directories by default; use --unsafe-output to bypass for one command.
• Secrets redaction — use --redact-secrets to mask common key/token patterns in --history and --export output.
• Network guardrails — provider base_url values are validated (https by default for non-local hosts), with configurable host allowlists, timeouts, and retry limits.
• Outbound warning guardrail — bot warns before printing responses that appear to contain dangerous shell command suggestions.
• Retention controls — use --purge to delete history, usage, and session files older than the configured retention window.
• Security log — export, clear, purge, and config changes are recorded in a local redacted JSONL audit log at ~/.bot/security.log.
• Security policy — see SECURITY.md for reporting guidance and supported-version details.
• Threat model — see THREAT_MODEL.md for threat boundaries, mitigations, and residual risks.
• API keys — keys are never written to disk; they are read only from environment variables at runtime.

---

Requirements

• Python 3.12+
• anthropic — Anthropic/Claude provider
• openai — OpenAI and llama.cpp providers
• rich — terminal rendering
• click — CLI framework

---

Cross-platform

Works on Linux, macOS, and Windows. Config and history paths resolve correctly on all platforms via Path.home(). File permission hardening applies on Unix only; Windows uses its own ACL-based access controls.

---

Built by

Inzen — AI consulting and LLM application development.

---

Contributing

git clone https://github.com/benwalkerai/bot
cd bot
uv sync --group dev
uv run pytest          # run tests
uv run ruff check .    # lint

CI runs automatically on every push and pull request via GitHub Actions — tests and ruff must pass before merging.

---

Releasing to PyPI

1. Update the changelog from conventional commits:

   uv run git-cliff --output CHANGELOG.md
   git add CHANGELOG.md
   git commit -m "docs: update changelog"
   

2. Bump the version — updates pyproject.toml, creates a commit, and tags it automatically:

   # patch: 0.2.1 → 0.2.2  |  minor: 0.2.1 → 0.3.0  |  major: 0.2.1 → 1.0.0
   uv run bump-my-version bump patch
   

3. Push the commit and tag — the tag triggers the PyPI publish workflow:

   git push && git push --tags
   

The publish.yml GitHub Actions workflow runs the tests, builds the package, and publishes to PyPI automatically via the PYPI_API_TOKEN secret.

---

License

GPL-3.0