Repository navigation
Monastery v0.1.2-alpha
Pre-releaseRelease Date: 2026-06-26
The second alpha. If v0.1.0 gave Monastery a brain — LLMs, chat, projects, Git — then v0.1.2 gives
it hands: it can now deploy what you build to your own infrastructure, dispatch agentic work to
Hermes, and read its own deployment failures and try to fix them. Still very much alpha: the
deployment paths depend heavily on your own homelab setup and you'll occasionally coach the LLM
through a build — but the core loop of prompt → build → ship to your own boxes now runs end to end.
🚀 Self-Hosting & Deployment (new)
The headline of this release: a Self-Host Wizard that takes a Monastery project, commits it to
your Git forge, and builds + deploys it on your own platform.
Coolify
- One-click deploy via the Coolify API (Dockerfile build pack)
- Clone-at-build Dockerfiles — the image clones your repo at build time, so deploys work with
IP /.local/ self-signed hosts that Coolify's normal Git validation rejects - Smarter server selection — de-prioritizes the built-in localhost server so apps land on your real
VPS, with a graceful fallback when a project has no deployment server - Deployment persistence (tracks the app UUID) for redeploy-in-place and 404 self-heal
- Base64 Dockerfile payloads and SSH/Git URL parsing fixes
Dokploy
- Full git-based deploy over Dokploy's tRPC API (create → save git provider → build type → deploy)
- Container port auto-detected from the Dockerfile
EXPOSEand applied before deploy - Fixes for auth, endpoint URLs, the dashboard link, the
saveGitProviderschema, and the
committed-Dockerfile flow (resolves "Github Provider not found" / "select-a-container")
Cloudflare Tunnel (both platforms)
- Optional auto-wired
cloudflaredsidecar — paste a tunnel token and Monastery deploys the
connector for you - Apps are published on a stable high host port for LAN + tunnel reachability; point your Cloudflare
Public Hostname at the returnedhttp://127.0.0.1:<port>service URL
Pocketbase backend (initial)
- Configure one shared Pocketbase instance in Settings; the wizard injects
POCKETBASE_URL
(build-time and runtime) into deployed apps - A toggle-gated instruction set teaches the connected LLM how to build against Pocketbase
(JS SDK usage, auth, collections) only when you're building a database-backed app
🔧 Deployment Monitoring & AI Auto-Fix (new)
- Pull a failed deployment's build log straight into chat with one click
- The connected LLM diagnoses the failure and applies fixes to your project files — then redeploy
- Graceful fallback: when a platform can't serve the remote build log (a known upstream Dokploy
limitation, issue #4687), the LLM proactively reviews your Dockerfile/config instead of dead-ending
🤖 AI Agents — Hermes Integration (new)
- Connect an external Hermes agent in Settings → Hermes Agent (proxied via
POST /api/hermes/run) - Agent mode toggle routes your chat through the agent instead of the raw LLM
- Agent roles — Architect, Coder, Reviewer, Tester, Documenter, Deployer — as a selectable lens
over your messages, multi-select capped at 2 - Hardened tool-call handling that surfaces Hermes's tool steps inline
- Shared-workspace docs so Hermes's own file writes show up in your Monastery project
⚡ Smarter Streaming
- Auto-continue — responses cut off by the model's output-token cap resume automatically
(capped, abort-aware, and toggleable), so long generations finish without manual clicking - Token usage shown per response when the endpoint reports it, accumulated across continuations
- Provider-aware response handling and token caps (local Ollama vs cloud) to stop premature cut-offs
- Fixes for runaway continuation loops and SSE stream parsing
🗂️ Workspace & Editor
- Create and delete files and directories directly from the sidebar
- File/folder uploads and drag-and-drop in the file tree
- SVG rendering in the preview pane
🎨 UI & Polish
- Consolidated chat toolbar — Auto-continue / Pocketbase / Agent-mode toggles collapsed onto a
single compact row, with agent roles revealed inline when Agent mode is on - Refreshed branding and LLM "thinking" animation
- Numerous compile and stability fixes across the Rust backend and React UI
local build servers, tunnels, registries), and you should expect to occasionally guide the LLM
through a build fix. Back up anything you care about before pointing Monastery at it.
Full Changelog: v0.1.0-alpha...v0.1.2-alpha