Stop offloading to AI,
Take command.
naholo runs your AI coding through one disciplined loop of brief → plan → ship → debrief. Every phase is chopped into review windows small enough to actually read so you stay in command while the agent does the chores.
The receipts back it up.
to review each checkpoint. Every window is sized to hold in one sitting, so you never drown in diffs.
fewer tokens burned. Throttled planning per phase means less wasted context.
throughput. One operator shipped a 5-day task in 1.5 days.
* Early results are anecdotal. The < 5 min review window is a design guarantee.
One cycle. Every op.
Six small moves from a rough idea to a shipped, reviewed change. The agent moves fast between them; you review at every gate.
- 1Drop the intel
/infil/fobLeave a rough idea or a pain point in the web app, or start straight from your codebase.
- 2Brief the direction
/warnoreview < 5 minThe agent pins the conceptual solution and the architecture decisions. It writes no tasks and no code yet, just a compact bullet list.
- 3Cut the plan
/opordreview < 5 min eachThe direction is chopped into single-commit-sized tasks. There is still no code, only a description of how the interfaces and module flows change.
- 4Ship one task
/splashreview < 5 min eachThe agent ships exactly one task. Anything it improvised is flagged as a deviation for you to accept or roll back.
- 5Continue the cycle
Review the next task and
/splash, or re-run/opordto adjust the plan mid-flight. - 6Debrief the spend
/exfilIt records the operation and its token usage to the server so you can look back. Weak plans cost more tokens, and now you can see exactly that.
Three reviews carry every change.
Agents write more than any human can keep up with, so you either vibe-code and ship slop or burn out trying to read it all. naholo throttles that output into a single document, built in small, swift steps that you review section by section.
You review a handful of decisions before any plan is written. It takes a couple of minutes, and because there is no giant document, no tokens go to waste.
Work is cut into single-commit tasks, each speccing how a module's interfaces and internal flow change. Staying at the contract level, not code, means fewer tokens to iterate and less time to review.
Each task ships in isolation, and any improvisation is flagged as a deviation. You confirm it before you trust it in front of anyone else.
Plan in the browser. Ship from the terminal/IDE.
Two tools, one loop: a web app that holds the context around every operation, and a CLI your agents drive to do the work.
Keep every op's context in one place
Think GitHub Issues, but for AI coding. Drop intel as a note or a quick message, and your agents pull the full context on /infil and report back on /exfil.
*Messages aren't an AI-chat wrapper. Use them for short memos and team communication.
#222 Redact ai transcripts
Provide skills/tools/commands for human and agents
Setup takes one step. naholo init connects the web app and installs the skills. The CLI ships optimized commands and an MCP server for your agents, keeping slop out and token usage low.
The skills walk agents through the workflow, so there's nothing to learn up front.
See exactly where tokens go
The CLI collects every transcript during an op and uploads it on /exfil. Transcripts are redacted by default, so everything but the usage data is purged, and you can opt out or send the full transcript at any time.
Usage is broken down by skill and mode, and cost is figured at Claude API rates rather than subsidized plan rates.
This is how you're supposed to do AI coding.
Turn vibe coding back into real engineering. Start your first operation, or fork it and self-host. It's MIT open source.
View on GitHub