Why it mattersA team can now keep one authoritative copy of every skill their agents use, revoke a leaked key without touching the skills themselves, and see which client actually loaded which revision.
Kitze, an indie developer whose earlier open-source work includes Sizzy and React Hanger, published Skillbox on 17 September. It is a self-hosted server that stores skills for AI coding agents as immutable revisions, hands them out over an MCP endpoint or a stdio bridge, and controls access through per-client keys that can be revoked one at a time. The repository is MIT licensed and reached 161 stars in the two days after its first push.
The problem Kitze is naming is the same one the earlier tool Kitter picked up on: a skill file copied into every project drifts, and nobody notices until an agent behaves differently in two repositories for reasons the diff cannot show. Skillbox answers it at a different layer. Instead of one local copy linked into projects, one central server owns the skills, and every client asks for them at run time.
How an agent uses it
An agent authenticates against the server's /mcp endpoint with a bearer token issued for that specific client. The base tools are search_skills, recommend_skills, load_skill, read_skill_file and report_skill_use. Skillbox also ships a Node and Bun stdio bridge for clients that do not speak HTTP MCP, and an install script that can configure Codex, Claude Code or Cursor against a given set of credentials.
A skill returned to an agent carries an immutable referenceId, the pinned revision, a description and a permission scope. The server never runs any skill code, and fetching a skill validates every path, file hash, size and package checksum before writing the files atomically. Revoking a client key blocks any future load. Files the client already downloaded stay on disk, and the README says so clearly.
Versioning, and who can change what
Every write produces a new revision. Skillbox detects conflicts on update and supports restore to an earlier version. Profiles carry independent create, update, archive and proposal permissions, so a shared installation can let an agent propose a skill change without letting it silently overwrite the master.
The optional recommendation layer
Skillbox has an optional recommend_skills mode that ranks the authorised skill catalogue against a task description. The ranking runs against either Vercel AI Gateway or the TypeSafe AI System One model, whichever the operator supplies a key for. The relevance score is described in the README as "an uncalibrated 0 to 4 rubric score, not probability", results below 3 are dropped, and any missing key, timeout or provider failure falls back to Postgres text search with an explicit fallbackReason. Skillbox does not ship a hosted account or provider credentials of its own.
The tech stack is React, Bun, Hono and PostgreSQL 16 or newer. Deployment is Docker-only with an optional Caddy HTTPS layer and Umbrel package files. Kitze notes on the release page that Umbrel packaging exists but has not yet passed real lifecycle verification.
For a team whose agents now run in more than one project, Skillbox moves the skills into one server that every agent asks over the network, and puts the read and write controls in one place instead of scattered across each repository.
Source
Primary source: kitze/skillbox on GitHub.
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.