Power Mode
Power Mode is Neural Inverse's agentic coding engine. It runs three specialized sub-agents in parallel — each with a scoped set of tools — to plan, implement, and verify code changes autonomously.
Power Mode is available in the open source edition. No Enterprise license required.
How It Works
When you start a Power Mode session, the parent agent spawns up to three sub-agents concurrently. Each sub-agent has a defined role with a strict tool whitelist — it cannot use tools outside its scope.
Sub-Agent Roles
| Role | Can do | Cannot do |
|---|---|---|
| Explorer | Read files, list directories, search paths, search file contents, read lint errors | Write or edit any file, run terminal commands |
| Editor | Everything Explorer can do, plus edit files, rewrite files, multi-replace content, create files/folders | Run terminal commands |
| Verifier | Read files, search, run terminal commands, run persistent commands, open terminals, send input, kill terminals | Write or edit files directly |
The Verifier closes the feedback loop — it runs your tests, reads lint output, and confirms the Editor's changes actually work before the session completes.
Additional roles available for specialized tasks:
- Debugger — read + search + terminal + edit, focused on hunting bugs
- Reviewer — read-only + search, for code review without any writes
- Tester — read + write + terminal, for writing and running tests
- Documenter — read + write + edit, for generating documentation
- Architect — read + search + agent research, for system design tasks
Session Lifecycle
Each Power Mode session has:
- A unique
idandtitle - An assigned
agentId(e.g.build,plan, or a custom agent) - A working
directory - A
status:idle|busy|error|compact - A full message history with role, parts, token usage, and cost
- A session
summary: lines added, lines deleted, files changed
Sessions persist across IDE restarts.
Tool Execution
Each tool call is tracked with:
status:pending|running|completed|errorinputparametersoutputorerrorresulttime.startandtime.endtimestamps- Optional
metadataandtitlefor display
Tools that require approval (configured via ask permission) will pause and prompt you before executing.
Built-in Tools
Power Mode includes 7 core tools available to all sessions:
| Tool | What it does |
|---|---|
bash | Run shell commands in the working directory |
read | Read file contents |
write | Write or overwrite a file |
edit | Make targeted edits to a file (search/replace blocks) |
glob | Find files by pattern |
grep | Search file contents by regex or string |
list | List directory contents |
Additional tools from other Neural Inverse features (firmware tools, GRC tools, browser tools, discovery tools) are registered when those environments are active.
Permissions
Each agent session can configure tool permissions per tool ID:
allow— tool runs without promptingask— tool pauses and asks for approval before runningdeny— tool is excluded from the session entirely
A wildcard * sets the default for any tool not explicitly listed.
Starting a Session
Open the Power Mode panel from the sidebar (or press the Power Mode keyboard shortcut). Choose an agent, set the working directory, and describe your task. The session begins immediately.
You can cancel any running sub-agent individually, or cancel all sub-agents at once.
Agent Reasoning
Power Mode messages include a reasoning part type — when your LLM supports extended thinking (e.g. Claude with extended thinking, o1/o3), the reasoning trace is captured and displayed inline in the session.
Maximum Concurrency
By default, Power Mode runs up to MAX_CONCURRENT_SUB_AGENTS sub-agents in parallel. Additional spawned sub-agents queue and start as running ones complete.