About Nightshift
Configure once. Sandbox everything.
Agent Environment
A self-contained world
Every Nightshift agent gets its own isolated directory at your chosen prefix. Tools, runtimes, config, and workspace — all provisioned at boot, all confined by the sandbox.
- Pinned toolchain — Python, uv, ripgrep, and opencode at exact versions
- Isolated XDG directories for config, cache, data, and state
- Scaffolded workspace with src, tests, .venv, and pyproject.toml
- AGENTS.md and SKILLS.md generated from BOOT.md at first run
- Sandboxed with bubblewrap — read-only host, writable workspace
Boot Loader
One file to rule the agent
Define your agent's entire environment in a single BOOT.md file. Skills, data schemas, tools, and behavior — all declared upfront and built by your agent at startup.
~/BOOT.md
1I need an agent for purchasing analytics. It should analyze procurement data,2track supplier performance, and generate inventory forecasts from our ERP data.34{5 "skills": [6 {7 "name": "data-schema",8 "description": "Schema reference for ERP tables"9 },10 {11 "name": "purchasing",12 "description": "Procurement optimization guidance"13 },14 {15 "name": "analytics",16 "description": "DuckDB patterns for parquet queries"17 }18 ]19}workspace/AGENTS.md
1# Purchasing Analytics Agent2Analyze ERP procurement data for spend insights, supplier performance, and inventory optimization.34## Skills5| Name | When to Use |6|------|-------------|7| data-schema | Before writing any query — check column names, types, relationships |8| purchasing | Supplier analysis, spend breakdowns, lead-time calculations |9| analytics | DuckDB syntax, parquet glob patterns, window functions |1011## Commands12```bash13uv run pytest # Run tests14uv run python -m procurement # Run analysis15uv run black . && uv run ruff check16```1718## Project Structure19src/procurement/20 loader.py — parquet ingestion with glob patterns21 suppliers.py — lead-time scoring, reliability metrics22 inventory.py — reorder points, safety stock23 metrics.py — velocity, EOQ, days-of-supply2425## Safety Boundaries26ALWAYS: Read files, run tests, generate reports to ./output/27ASK FIRST: Install packages, modify pyproject.toml, create exports28NEVER: Delete ./data/*, run rm -rf, call external APIsAgent Sandboxing
Read-only host. Writable workspace.
Nightshift uses bubblewrap to mount the entire host filesystem as read-only. The agent can only write to its own isolated environment. No accidental overwrites. No system damage.