One command, no shell
The shell connector execs one command with no shell: chains, pipes and redirections are unrepresentable. Every other tool is read-only — asserted by a test, not by trust.
A model writes your code, you decide what to keep. FastAPI behind, Vue 3 in front, one browser tab in between — and it puts that agent to work on a real checkout, in two pages.
Not · Just · Another · Editor
Two pages, one folder
Both tabs are the same workspace. One drives the conversation, the other shows the tree the conversation is changing — so nothing the model does is hidden behind a diff you approve blind.
A streaming chat with the agent you pick. Chat has no tools at all; Code gets the file tools and the language servers; Architect reads your project and writes a spec, never your code.
An agent replaces the persona sent to the model, names the model the turn runs on, and decides whether a whole tool hub reaches it at all — and it can only ever narrow what Settings already allows.
This tab's folder as a file tree and a Monaco pane: create, rename, move and delete, save with a diff against the last commit, git status badges on the rows, and one terminal per connection on the platform's pseudo-terminal.
The code pane is served by the same language servers the WorkBench tools use — Ctrl+click a name for its definition, hover for its type, and real diagnostics in the status bar.
Answers, reasoning, tool calls and timings arrive over Server-Sent Events — the UI's only source of truth. When your endpoint answers ask again later, nojae waits out the pause and re-sends the identical request, so a provider outage resumes instead of ending a run.
What you get
Your code and your endpoint stay on your machine. The agent's reach is bounded by a per-workspace read/write policy — and the connectors answer through it, so a folder you mark read-only is not editable by an agent either.
The shell connector execs one command with no shell: chains, pipes and redirections are unrepresentable. Every other tool is read-only — asserted by a test, not by trust.
Works through a folder of task files unattended. A task is finished when its file leaves the lot folder — that move is the agent's own acknowledgement, and a file still there is relaunched up to the budget.
Every save keeps the bytes it replaced in a backup, and a delete moves
the file into a dated batch under .nojae/trash/. A restore
is one mv away.
Connectors and language servers are declared per workspace in a file, hot-toggled from the UI, and mirrored by two hubs that share one path policy with the Editor.
Drop a SKILL.md folder in .agents/skills and
the agent starts every conversation knowing it exists — name and
description only. The body is read on the model's own judgement.
Session clock, generation speed and context occupancy, measured app-side on chunk arrivals — so tool execution never enters the numbers, and a 40-minute outage is not a 40-minute session.
Up close
Click any screenshot to enlarge it.
Quick start
You need Python ≥ 3.10 and Node.js. Desktop browsers only.
# install
python -m venv .venv && source .venv/bin/activate
pip install -e ".[mcp,lsp]"
cd web && npm install && npm run build && cd ..
# run
nojae
# → open http://127.0.0.1:8001
mcp brings the WorkBench connectors, lsp
the language-server tools, windows the terminal's
pseudo-console driver. None is mandatory.
One command, and it serves the built front end on
127.0.0.1:8001.
Settings → Servers: your OpenAI-compatible endpoint —
http://127.0.0.1:8080/v1 for a local llama.cpp —
press Fetch models, add a model, give it to an agent.
A fresh install has no model, and no turn can run until you give it one. Nothing else is required for the Editor; the connectors only decide how far the agent may reach.
Apache 2.0, self-hosted, and the whole thing is one FastAPI app in front of one Vue 3 build.
« On ne dit kenavo qu'à ce qui peut nijañ. »