# nebu-ctx — Context Engineering Layer (Global)
You have the nebu-ctx MCP server available. You MUST use it for ALL file reads, shell commands, and code searches. Using native tools wastes tokens — nebu-ctx compresses everything.
## Tool Replacement Rules
| `Read` / `View` / `cat` / `head` / `tail` | `ctx_read(path, mode)` — cached, 10 read modes, re-reads ~13 tokens |
| `Bash` (any shell command) | `ctx_shell(command)` — pattern compression for git/npm/cargo/docker |
| `Grep` / `Search` / `rg` | `ctx_search(pattern, path)` — compact, token-efficient results |
| `ListFiles` / `ListDirectory` / `ls` / `find` | `ctx_tree(path, depth)` — compact directory maps |
## How to Use
```
ctx_read("src/main.rs") # instead of Read("src/main.rs")
ctx_read("src/lib.rs", mode="map") # deps + API signatures
ctx_read("src/lib.rs", mode="signatures") # API surface only
ctx_shell("git status") # instead of Bash("git status")
ctx_search("pub fn", "src/") # instead of Grep("pub fn", "src/")
ctx_tree(".", 2) # instead of ListFiles(".")
```
Write, Edit, and other mutation tools have no nebu-ctx equivalent — use them normally.
CRITICAL: Every time you reach for Read, Bash, Grep, or ListFiles — stop and use the lean-ctx MCP equivalent instead. This is not optional.
## Memory policy
- Use `ctx_session` for task state, findings, and decisions that should survive across chats.
- Use `ctx_knowledge(action="remember")` for durable project facts.
- The stop/compact hooks already persist session state into the nebu-ctx server; rely on that instead of chat history for reusable memory.
Preferred workflow: use `ctx_workflow` for state + evidence + tool gating.