Hexis

[Hexis by Bevel: git-backed skills, tools and context for AI agents]

[GitHub stars] [Live demo] [Apache-2.0 license] [Last commit]

Git-backed control plane for AI-agent skills, tools, context, permissions and identity. Self-hosted and MCP-native.

One place where your company's AI plugins, tools and knowledge live: centrally managed, reviewed and access-controlled, and usable from any AI agent. The open-source core of the Bevel platform.

Why Hexis?

Used by teams at

[osapiens] [BEO] [Unite] [Workpath]

Read more about the learnings that we made which led to Hexis here: Our Medium Article

For teams

One place where engineers and non-technical people alike can browse and load plugins, propose suggestions, and manage access.

For enterprises

Every skill, tool manual and permission is a file in a git repository you own, so the audit trail is the storage layer: who changed what, when, who approved it, and how to undo it. An agent can only do what the person running it can do, resolved per file, and it never holds the credentials it uses. Runs on your infrastructure, behind your own SSO.

Why it's different from MCP gateways:

MCP gateways are uni-directional; users can consume plugins, skills or tools but there is no mechanism here for users to propose changes or share new skills and MCP servers. You can do this via GitHub in the back, but this is not accessible to non-technical users and there is no fine-grained access control for either viewing or the review process.

Hexis can do all of the above specified capabilities for distribution, and has this bidirectionality needed for management.

In Hexis, skills and tool manuals are reviewable files. Anyone can propose a change; on protected branches it reaches the owners of the files it touches and ships only once they approve. Agents propose too: one that hits a broken skill mid-task can suggest the fix, and a person decides whether it lands.

Contents

• See Hexis in action
• Connect Hexis to Cline
• Try the live demo
• Managed hosting
• Deploy with Docker
• Links and images in knowledge pages
• Local development
• Configuration reference
• Skills in Cowork and claude.ai
• Troubleshooting
• Repository layout
• FAQ

[Watch the full Hexis demo]

Watch the full walkthrough: connect an agent, use company context, review proposed changes, and manage team access.

See Hexis in action

Propose and approve skill changes

Anyone can propose a new skill or improve an existing one. On protected branches, owners review the exact change and approve it before it becomes available to the team's agents.

[A teammate proposes a skill and its owner reviews and approves the change request]

Use your team's skills in Claude

Connect Claude to Hexis over MCP, then ask normally. Claude can discover and load the approved skill instructions and company context your role can access, without copying prompts between tools.

[Claude uses approved team skills and company context through Hexis over MCP]

Connect Hexis to Cline

Cline can connect directly to Hexis as a remote Streamable HTTP MCP server. Install the public demo connection from the Cline CLI:

cline mcp install hexis --transport http https://demo.bevel.software/api/mcp --yes

Complete the OAuth sign-in in your browser when prompted. Cline then discovers the skills, tools and context your Hexis role can access. For your own Hexis deployment, replace demo.bevel.software with your deployment's host.

Share skills with the right people

Add teammates to roles or grant access directly when needed. Everyone connects to the same workspace, while each person and their agent only sees what they are allowed to read.

[An owner adds teammates to roles and manages access to shared company content]

Try it first: the live demo

demo.bevel.software is a public instance you can sign into with your Google account, populated with a fictional company's knowledge, skills and tools. The Start here page walks you through the whole loop: connect your own agent over MCP, have it build a sales deck from a skill, watch its proposed improvement arrive as a change request. The demo is shared and read-mostly (visitors propose, owners approve); everything below gets you the same thing with none of the limits.

Want a managed instance?

We run it for you (hosting, upgrades, backups, SSO) and your team just signs in. Write to [email protected].

Deploy it in 5 minutes (Docker)

You need: Docker with Compose on a server (or your laptop; one extra line below), and an empty git repository on any host (GitHub, GitLab, Bitbucket, Azure DevOps, self-hosted) to hold your knowledge base. The app seeds it with a starter template on first run.

Grab the two deployment files — no clone needed:

mkdir hexis && cd hexis# v0.15.1 below = the release this page was written against; replace with the latest release tagwget https://raw.githubusercontent.com/Bevel-Software/Hexis/v0.15.1/docker-compose.ymlwget -O .env https://raw.githubusercontent.com/Bevel-Software/Hexis/v0.15.1/.env.example

(Working from a git clone works identically — both files sit at the repo root; cp .env.example .env.)

Open .env and fill in the four required values (everything else can wait):

[email protected]     # the deployment owner, always an adminADMIN_PASSWORD=pick-something   # sign-in password; only with password login (SSO-only deployments drop it)JWT_SECRET=…                    # generate with the command belowSECRETS_ENC_KEY=…               # generate with the command below

Generate the two secrets (run twice, paste one result into each):

node -e "console.log(require('crypto').randomBytes(32).toString('base64'))"# no Node installed? docker run --rm node:22-slim node -e "console.log(require('crypto').randomBytes(32).toString('base64'))"

For a public deployment served over HTTPS by the bundled proxy, also set the domain — it derives everything else public (origins, proxy hop):

DOMAIN=bevel.your-domain.com

Then start everything. Deploying pulls the image CI publishes on every release — nothing compiles on your server, so a small instance suffices. Pin the version in .env (HEXIS_VERSION=0.15.1) so a later pull can't become an unplanned upgrade — UPGRADING.md covers upgrades and backups. Building from source instead (a staging server tracking a branch, a fork) is deployment/docker-compose.build.yml (explained in deployment/).

Public HTTPS, no proxy of your own (a bare EC2 instance, a plain VPS): the https profile starts Caddy in front of the app, with automatic Let's Encrypt certificates for DOMAIN and the HTTP→HTTPS redirect. First: a DNS A (or AAAA) record for the domain pointing at the server, and ports 80 + 443 open to the internet (port 80 is not optional — the certificate challenge and the redirect both use it).

docker compose -f docker-compose.yml --profile https up -d

Behind your own reverse proxy (Coolify, Traefik, nginx): skip the profile — two things terminating TLS for one app is one too many. Instead of DOMAIN, set the origin values and the proxy hop count in .env, so OAuth redirects are built right and rate limits see real client IPs instead of the proxy's:

PUBLIC_BACKEND_URL=https://bevel.your-domain.com   # public origin; OAuth redirects are built from itPUBLIC_FRONTEND_URL=https://bevel.your-domain.com  # same origin: the backend serves the SPATRUST_PROXY=1                                      # your proxy hop count

docker compose -f docker-compose.yml up -d

The explicit -f matters in a clone: it skips docker-compose.override.yml, so the app publishes no host port — your proxy reaches it on port 3001 over the compose network. This is deliberate: a fixed published port makes every redeploy fail with port is already allocated, because the replacement container starts while the outgoing one still holds it.

Open your domain and sign in with ADMIN_EMAIL / ADMIN_PASSWORD.

Just trying it on your laptop? Same steps, minus DOMAIN and the origin values — plus one extra file: docker-compose.yml alone publishes no host port (see above), and docker-compose.override.yml is the piece that puts the app on localhost. A clone already has it; next to the wget'd files, fetch it too:

wget https://raw.githubusercontent.com/Bevel-Software/Hexis/v0.15.1/docker-compose.override.ymldocker compose up -d

Then open http://localhost:3001 (a different port: APP_PORT=8080 docker compose up -d). Leave TRUST_PROXY unset here — with no proxy in front, trusting forwarded headers would let clients spoof their own address.

First sign-in: the setup screen

The app asks for the things it could not guess, and tests them against the real host before saving:

1. Knowledge-base repo: the https clone URL of that empty repository.
2. Git credential: a token with read/write access to it (for GitHub: a fine-grained personal access token with Contents: read & write on that one repo is enough).
3. Branch model: which branch is the default and which are protected (changes to protected branches only land through approved change requests). The repository's real branches are offered as suggestions; for an empty repo the default (main) is fine.

Since the repo is empty, the app initialises it from the bundled template and writes a roles.yaml whose first Admin is you. That's it: you're in the workspace. Head to Skills & Tools to make your first plugin and skill, and to Connect (in the app menu) to hook up an agent over MCP.

Going to production? Configuration reference covers single sign-on, the state you need to back up, health checks, and configuring by environment instead of the setup screen.

Links and images in knowledge pages

Pages are markdown. A link to another page is a relative path, and Hexis opens it in the app:

See the [approval process](../Processes/Approval.md#steps).

Images work the same way. Keep them in an assets/ folder next to the pages that use them, and link them relatively:

![Approval screen](./assets/approval-screen.png)

Access follows folders, so a person who may read the page may see its screenshots, and moving the folder keeps every link valid; one shared Uploads/ folder gives up both. An export that arrives with a sibling .assets/ folder (Microsoft Loop, for one) can be dropped into the knowledge base as it is, and the links resolve unchanged. Pasted base64 images are not supported: save the file and link it.

Local development (run from source)

You need: Node 22.13 or newer (.nvmrc; the engine range is >=22.13 <23), pnpm 10, git ≥ 2.41, ripgrep (rg on your PATH — the agents' grep tool searches with it, and the server and the backend test suites refuse to run without it), and a Postgres 17 (the bundled one is fine):

docker compose up -d db        # just the databasepnpm installpnpm build                     # builds the packages the apps importcp .env.example .env           # fill the same four required values;                               # the default DATABASE_URL already points at the bundled dbpnpm dev                       # backend on :3001, Vite dev server on :5173

Open http://localhost:5173 (the dev server proxies to the backend). Useful commands: pnpm test, pnpm typecheck, pnpm lint.

Migrations run automatically on boot; there is no separate migrate step, in dev or in production.

Reference

• Configuration: every environment variable, SSO setup, secret generation, backups and health.
• Several knowledge bases in one process: the tenants file, what is kept apart, and moving a tenant to its own deployment.
• Troubleshooting: the failures you are most likely to hit, and what causes them.

Repository layout

FAQ

Questions that come up when teams evaluate Hexis as a central, versioned catalogue for agent skills and tools.

License: Apache-2.0