Skip to main content

​
Install

curl -sSfL https://xynthis.com/install.sh | bash
The installer detects your platform, downloads the matching release tarball, verifies its SHA-256, installs the binaries into ~/.xynthis/bin/, registers the brain daemon to start at login (launchd on macOS, a systemd user unit on Linux), starts it, and adds ~/.xynthis/bin to your PATH. Total time is about 5 seconds plus the download. On Windows, use PowerShell instead: irm https://xynthis.com/install.ps1 | iex. See Installation for what differs there. Open a new shell so the PATH update takes effect.

​
Verify

xynthis health
health checks the whole stack: binaries, config, and the brain daemon behind its socket. For the brain’s live counters: 
xynthis status
If something is off, xynthis doctor diagnoses config, auth, and binary issues.

​
Pick a model

Xynthis ships 14 built-in provider presets. See what’s available and wire one up: 
xynthis models list --all
xynthis login anthropic          # interactive API key entry
xynthis models set anthropic
No API key? Use a local model via Ollama (xynthis models add ollama, then xynthis models set ollama), or reuse an existing Claude Code session with xynthis models set claude-code. xynthis auth scan finds credentials already on your machine.

​
First conversation

xynthis chat "hello, do you remember me?"
chat runs one agent turn against the active model, prints the reply, and exits. Follow-up chat calls append to the same session, so the conversation has continuity; pass --new to start fresh. Running bare xynthis opens an interactive REPL instead.
On macOS, chat launches the menu-bar app alongside so you can watch the agent work. Pass --no-app (or set XYNTHIS_NO_APP=1) to stay in the terminal.

​
Teach it something

xynthis remember "the staging database password rotates every Tuesday"
remember stores a free-form note directly into the brain. It also learns passively: facts from your conversations land in memory without an explicit remember.

​
Get it back

Two ways, depending on whether you want raw matches or an answer: 
xynthis recall "database password rotation"
recall hits the brain’s layered retriever and prints the matching records with scores. It bypasses the agent entirely. 
xynthis ask "when does the staging password rotate?"
ask answers factoid questions straight from memory: no LLM call, zero tokens, milliseconds. When no match clears the confidence threshold (0.5 by default, tune with --threshold), it still shows what it found and suggests xynthis chat for a reasoned answer.

​
Where your data lives

Everything stays in ~/.xynthis/ on your machine: 
~/.xynthis/
├── bin/            binaries
├── bmc/            binary memory stores (persistent)
├── sessions/       per-session chat history + working-memory ledgers
├── skills/         installed skill library
├── lessons.json    the agent's failure→correction playbook
├── config.toml     model + provider config
├── auth.json       API keys (mode 0600)
├── logs/           daemon stdout/stderr
└── brain.sock      unix socket — the brain's only API
Nothing here is sent anywhere on its own. When you use a remote LLM provider, the request carries your message plus the query-relevant memory the agent recalls and injects into the prompt; the store itself is never uploaded.

​
Next steps

  • Installation: what the installer does in detail, Windows, upgrading, uninstalling.
  • The macOS app: the menu-bar resident client.
  • Models: configuring providers, local models, and switching.