infai
██╗███╗ ██╗███████╗ █████╗ ██╗
██║████╗ ██║██╔════╝██╔══██╗██║
██║██╔██╗ ██║█████╗ ███████║██║
██║██║╚██╗██║██╔══╝ ██╔══██║██║
██║██║ ╚████║██║ ██║ ██║██║
╚═╝╚═╝ ╚═══╝╚═╝ ╚═╝ ╚═╝╚═╝
Two terminal tools for local-first AI, from one repository.
Each is a single Go binary for macOS and Linux, released and installed separately. Use either one on its own, or both together.
infai: local inference
Every inference engine has its own dialect of flags. infai gives them one shape, the profile: model + engine + config. Define it once, then launch it with a keypress.
Install
brew install dipankardas011/tap/infai # macOS and Linux
- Linux packages: install
infaifrom the openSUSE Software portal, which has repository setup for Debian, Ubuntu, Fedora, and openSUSE. - Script (Linux): installs to
/usr/local/bin.curl -sL https://raw.githubusercontent.com/dipankardas011/infai/main/install.sh | bash - Binary: download one from Releases (Linux and macOS, amd64 and arm64).
Quick start
infai # open the TUI
infai --version
On first launch infai creates its local database. In the Engines tab, add your
llama.cpp or vLLM binary. In the Models tab, add a model folder and press s to
sync it, or press w to download from Hugging Face. Then create a profile in
Profiles and press enter to launch it.
Features
- Profiles: named launch configurations (context size, GPU layers, KV-cache types, flash attention, sampling, ports, extra flags), with engine-aware options for llama.cpp and vLLM.
- Runs dashboard: run several models at once, with status, uptime, and live tokens per second for each one. Ports are checked before launch.
- Live telemetry and logs: system and per-model CPU, RAM, and VRAM, plus a scrollable server log you can copy.
- Model library: scans folders for GGUF and SafeTensors files, reads their metadata, and links multimodal projectors automatically.
- Hugging Face downloads: search the Hub, pick a GGUF quantization, and download with resumable, atomic transfers. Provenance stays attached to the model.
- Hardware-fit checks: estimates whether a model fits your hardware and suggests launch settings before you start.
- 11 themes: tokyonight, everforest, onedark, rosepine, gruvbox, catppuccin, nord, dracula, kanagawa, solarized, monokai.
Keys
tab / shift+tab switch tabs, t picks a theme, ? shows help, q quits.
Data
Profiles, folders, engines, and settings live in one SQLite file. Back it up with cp.
Schema migrations run automatically on startup.
infaiw: agent harness
infaiw runs agents as a background service, so a session keeps working after you
close the terminal. The infaiw TUI is a client: attach from several terminals,
leave, and come back to the same session.
- Sidecar loops: delegated subagents that are not done until an acceptance check passes. Each loop is its own session and is recorded on the timeline.
- Durable, branching timelines: every session is an event timeline that is compact to store and fast to replay, and you can branch from any point.
- Human first: approvals wait for you even after you close the window. Approve, deny, or deny with a reason.
- Inter-session messaging: sessions have mailboxes, and the engine routes messages between them.
- Task checklists with intent: each task records what was expected, so the agent can recover the goal mid-task.
- Hybrid auto-compaction: recent events stay verbatim and older ones are compressed.
- Tools: files, bash, web fetch and search, images in and out, skills, MCP servers (with OAuth, elicitation, and prompts), and evals.
- Providers: DeepSeek, OpenAI Codex, and Claude Code through your local login, with configurable thinking levels.
Install and start
The service is opt-in: installing the package does not start it. Run these as your
normal user, without sudo.
Homebrew (macOS and Linux)
brew install dipankardas011/tap/infaiw
brew services start infaiw
Linux packages: install infaiw from the
openSUSE Software portal, then:
systemctl --user daemon-reload
systemctl --user enable --now infaiw.service
Quick start
infaiw provider login # sign in to a model provider
infaiw # open the TUI and start a session
The TUI attaches to localhost:6000 by default. Use --host and --port to reach
another server, and --session <uuid> to resume a saved session.
Claude Code
This provider uses the Claude Code CLI already installed and signed in on the same
machine. There is no API key: infai never reads, stores or forwards the CLI's login.
It requires Claude Code 2.1.301 or newer, which you can check with claude --version;
adding the provider fails with an actionable message if the CLI is missing, too old,
or not signed in.
The systemd user manager does not read shell startup files, so its PATH does not
include ~/.local/bin where the native install lives. The packaged unit adds it, and
~/.config/infaiw/server.env can override the service environment when the CLI is
somewhere else.
infai keeps the agent loop, so the CLI runs with its own tools, MCP servers, settings and slash commands disabled, in an empty infai-owned directory, and without writing a session. Tool requests come back through a JSON Schema infai supplies, and infai executes them under its own approvals.
When the service delegates its cgroup — the packaged unit does, on systemd 254 or newer — each turn is capped at 500m CPU and 512 MiB of memory, entered at process start. A turn that exceeds the memory cap is reported as a provider failure naming the limit. Hosts without delegation run the provider unconstrained.
Keys
Running the service
Homebrew manages launchd on macOS and systemd on Linux. Use
brew services stop infaiw or brew services restart infaiw. Logs are in
$(brew --prefix)/var/log/infaiw.log. Upgrading the formula stops a running
infaiw server, and the service starts the new binary; an inactive service stays
inactive.
If you installed the old cask, migrate once:
brew uninstall --cask infaiw
brew update
brew install dipankardas011/tap/infaiw
brew services start infaiw
The executable moves from the Caskroom to the Cellar. Your user data is not moved or removed.
--userruns the service as your account, not root.enablestarts it on future logins;--nowalso starts it immediately.- Follow the logs with
journalctl --user -u infaiw.service -f(ctrl+cstops following, not the server). - Package upgrades restart running instances; inactive ones stay inactive.
- Stop it and disable startup with
systemctl --user disable --now infaiw.service. Your data and configuration are kept.
Without lingering, a user service stops after your last session ends. To keep it running after logout and start it at boot (may need admin rights):
sudo loginctl enable-linger "$USER"
Packaging follows the Homebrew service DSL,
openSUSE unit guidelines,
and Debian's dh_installsystemduser
with --no-enable.
Configuration
infaiw server is configured only through INFAI_AGENT_* environment variables, for
example INFAI_AGENT_PORT=6000. Background services do not inherit your shell
environment, so put them in ~/.config/infaiw/server.env (chmod 600 if it holds
secrets) and restart the service.
Building from source
Requires Go 1.26+. infai also needs a C compiler, because it uses SQLite through CGO; infaiw does not.
git clone https://github.com/dipankardas011/infai.git
cd infai
CGO_ENABLED=1 go build -o infai ./cmd/inference
CGO_ENABLED=0 go build -o infaiw ./cmd/agent
go test ./...
Contributing
Bug reports and pull requests are welcome on GitHub. See AGENTS.md for the architecture and conventions.
The website lives on the gh-pages branch; edit it by committing there. Provider
catalogs are the exception: they live in catalog/ on main, next to the code that
reads them.
infaiw is being built in public, one post a day; follow along from the build log.

