SYNAPTOME

Pre-release · open source · runs on your machine

One intelligence layer.
Every capability.

Synaptome finds skills, agents and MCP servers in public sources, audits each one as untrusted input before anything can run, installs only what you approve, and routes each task to the few that fit — so your AI coding host reads short summaries instead of everything you have installed.

InstallSource

The Web Command Center overview: capabilities, pending approvals, installs, discovery sources and recent activity.

How it works

Discover

Nine sources — GitHub, the official MCP registry, npm, PyPI, GitLab, official skill repositories, curated lists and the skills already on your machine — each read within its terms, backoff and daily budget.

Vet

Everything downloaded is quarantined and audited: prompt injection, hidden characters, pipe-to-shell, exfiltration, licences, permissions. Executable content always needs a person.

Install, with a way back

Into a sandbox copy of the host first. A real install needs explicit confirmation, never overwrites what Synaptome did not create, and can always be rolled back exactly.

Route

A connected project asks which capabilities fit a task and gets short summaries within a token budget — and nothing at all for trivial prompts.

Improve

The Skill Forge turns a gap into a tested skill and promotes it only on held-out evidence from an independent author.

Three interfaces, one state

A terminal workbench, a web command center on loopback, and a CLI with JSON output. A change in one appears in the others.

Safety, stated plainly

Install

Requires Node.js 24 or later.

npm install -g synaptome
synaptome init
synaptome            # Terminal Workbench
synaptome dashboard  # Web Command Center

Quick start in fifteen steps · Install, upgrade and uninstall

Status

Version 0.2.0 is a pre-release. Windows is the tested platform, including a clean install from npm; on macOS and Linux the test suite passes in CI, and a real install there is not yet verified. Executable capability content is never tested dynamically, so it always needs approval. Of the model providers, only headless Claude Code is verified by a live call. What works and what does not, with evidence: acceptance matrix.