infrctl

Five local AI model families. One CLI.

infrctl is a local-first command-line tool for running AI models from your terminal through Ollama. It keeps the user-facing model surface intentionally small: Qwen, DeepSeek, Llama, Gemma, and Phi.

CA: 0x22d0a50bb8b5789655ae1c84b34dc7804b0378e2
Dexscreener: https://dexscreener.com/base/0xfea8fb6981c153cfc66c855687290f441f94018db40627aa09c8fb79a631ce7b

Install

Standalone install, no Node.js required:

curl -fsSL https://raw.githubusercontent.com/infrctl/infrctl-cli/main/install.sh | sh

The installer downloads the right binary from GitHub Releases into ~/.local/bin, verifies the release checksum, and installs Ollama automatically if it is missing on Linux or macOS.

npm install, for Node.js users:

npm install -g infrctl

To inspect the installer first:

curl -fsSL https://raw.githubusercontent.com/infrctl/infrctl-cli/main/install.sh -o install.sh
sh install.sh

Requirements:

• Linux or macOS for the curl installer.
• curl or wget.

Node.js 18+ is required only for npm installs or local development.

If you already manage Ollama yourself, skip the bundled Ollama install:

curl -fsSL https://raw.githubusercontent.com/infrctl/infrctl-cli/main/install.sh \
  | INFRCTL_SKIP_OLLAMA=1 sh

Quickstart

infrctl setup --starter
infrctl
infrctl -p "say hello from infrctl"
infrctl smith "explain this repo"

The starter setup installs Ollama if needed, starts it when possible, and pulls a lighter first-run model set: Phi and Qwen. Add more models later:

infrctl pull deepseek
infrctl pull llama
infrctl pull gemma

Session-First CLI

infrctl is designed to feel like a modern interactive terminal assistant:

infrctl                         # start interactive chat
infrctl qwen                    # start chat with Qwen
infrctl -p "explain this log"   # print a one-shot answer
infrctl -p "continue" --continue
infrctl --resume <session-id>
infrctl sessions

Use JSON output for scripts:

infrctl -p "say hello" --output-format json

Pipe stdin into one-shot mode:

cat README.md | infrctl -p "summarize this"

Interactive chat sessions autosave locally:

~/.infrctl/sessions/
~/.infrctl/chats/

Inside chat:

/help
/status
/transcript
/switch qwen
/system <prompt>
/temp 0.4
/clear
/exit

Smith Coding Agent

Smith is the local coding agent inside infrctl:

infrctl smith "find the config loading code"
infrctl smith "add tests for the status command"
infrctl smith --model qwen "refactor this function"
infrctl smith --dry-run "show the patch you would make"
infrctl smith --json "explain this repo"

Run it without a task to open an interactive Smith session:

infrctl smith

Smith works inside your current directory by default. To point it at another repo:

infrctl smith --cwd /path/to/repo "summarize the command structure"

Safety defaults are intentionally conservative:

• Smith can scan, read, and search your repo.
• Smith drafts unified diff patches and asks before applying them.
• Smith asks before running shell commands.
• Smith saves reversible patch backups for undo.
• Smith reads key repo files like AGENTS.md, README.md, package.json, and framework config.
• Smith blocks destructive, publishing, deployment, and private-key-related commands in V1.

Permission profiles:

infrctl smith --profile safe "explain this code"       # read/search only
infrctl smith --profile normal "fix this test"         # default approvals
infrctl smith --profile fast "fix and run checks"      # auto-edit safe patches
infrctl smith --profile danger "move quickly"          # still blocks destructive commands

Approval modes:

infrctl smith --approval ask "add a small test"        # default: approve patches
infrctl smith --approval step "inspect this bug"       # approve each action
infrctl smith --approval auto-edit "fix a typo"        # apply safe validated patches

Shell modes:

infrctl smith --shell ask "run tests after the fix"    # default: ask first
infrctl smith --shell safe "fix and check the build"   # auto-run safe checks
infrctl smith --shell off "only propose file edits"    # never run commands

Undo the latest Smith-applied patch for the current workspace:

infrctl smith --undo

Inside interactive Smith:

/status
/diff
/apply
/reject
/undo
/run npm test
/test full
/model qwen
/compact
/exit

Smith sessions are saved locally:

~/.infrctl/agent-sessions/
~/.infrctl/agent-patches/
~/.infrctl/repo-memory/

For coding tasks, Qwen and DeepSeek are usually the strongest defaults. Phi is useful for tiny machines and quick edits.

Commands

infrctl setup
infrctl setup --starter
infrctl models
infrctl pull qwen
infrctl pull all
infrctl chat qwen
infrctl ask qwen "hello"
infrctl serve qwen
infrctl serve qwen --auto-port
infrctl status
infrctl smith "add tests for config loading"
infrctl doctor
infrctl config show
infrctl sessions
infrctl resume
infrctl completion zsh
infrctl update

If a selected model is missing, ask, chat, and serve offer to pull it. Use --yes to skip the prompt:

infrctl -p "hello" --yes

Supported Models

V1 exposes exactly five model families:

• Qwen
• DeepSeek
• Llama
• Gemma
• Phi

The real Ollama tags are selected through a local registry and hardware-aware recommendations.

Local API

Start the local API server:

infrctl serve qwen

If the default port is busy:

infrctl serve qwen --auto-port

Chat completions:

curl http://127.0.0.1:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen",
    "messages": [
      { "role": "user", "content": "Hello" }
    ]
  }'

Health check:

curl http://127.0.0.1:8787/health

Shell Completions

infrctl completion bash
infrctl completion zsh
infrctl completion fish

Troubleshooting

Check your environment:

infrctl doctor
infrctl status

If Ollama is missing:

infrctl setup

If Ollama is installed but not running:

ollama serve

If a server port is busy:

infrctl serve phi --auto-port

If a model is missing:

infrctl pull phi

Update

Standalone users can rerun the installer:

curl -fsSL https://raw.githubusercontent.com/infrctl/infrctl-cli/main/install.sh | sh

Node.js users can update with npm:

npm install -g infrctl@latest

From inside the CLI:

infrctl update

Privacy

infrctl runs locally through Ollama. It does not send prompts to cloud APIs. It does not collect telemetry.

Development

cd infrctl
npm ci
npm run check
npm run build:binary

Release checklist:

docs/release-checklist.md

License

MIT