AIWatcher Local
Private and localYour AI coding work, outcomes, and improvement signals.
Plan
Paste the prompt you are about to send. AIWatcher chooses the route before you spend context: rewrite, fork, Fresh Start, archive, or continue.
Use this when a surface does not reliably invoke hooks. Hook-capable tools can get the same local analysis automatically; this tab is the manual bridge. Nothing is sent anywhere or intercepted on your behalf, and prompt text is analyzed locally and not persisted.
Route
One recommended action first, evidence second.
Projects
Local repos and folders absorbing AI coding work.
| Project | Status | Sessions | Tokens | Model calls | API value |
|---|
Cost against size, one dot per session
Both axes are logarithmic, so a dearer model sits higher rather than climbing more steeply. Look for a hollow dot high up: an expensive session that produced nothing. A faded dot is one nobody has judged yet, which is not the same as one that failed. Click any dot to open that session.
Not a verdict on which model is better value. If the dear model gets the hard problems it will land less often for reasons that have nothing to do with the model, and nothing local separates those.
Models and Tools
Where the spend went, by model and by tool.
By model
By tool
Model mix within each tool
Each bar is that tool's own 100%, so size means proportion, not volume — the token total is on the right.
Context health
Context bloat, runway pressure, and Fresh Start actions for local work.
Sessions
Search sessions, inspect projects, and continue prior local AI work. Prompt text is shown for your own review only, never uploaded.
Every row says what AIWatcher knows: active app, likely workspace, historical log, user-marked outcome, or inferred local evidence.
Cost per change
What each commit cost in AI spend, and what that works out to per line.
A change's cost is the AI spend in that repo since the previous change, attributed per model call rather than per session, and capped at a 12h lookback from when the work was authored — rebasing rewrites a commit's date, and keying off that stranded the spend behind it. Survival is only measured for changes old enough to judge and costly enough to reach the sampling budget — a blank means not measured, not "did not survive". It is a floor either way: reformatting moves attribution away from the original change.
Optimize workspace
Unused AI sessions and scratch workspaces ready for review.
Was any of this worth it?
What the spend bought, what it did not, and how much of it can be judged yet.
Outcomes and guardrails
What stuck, and what was caught before it ran.
Fresh Start receipts
- preflight decisions · -. When AIWatcher suggested a fresh session, what you chose, and whether a follow-up session was observed. Proof stays pending until a later session in the same project is observed.
| Time | Decision | Expected context at risk | Proof | Sessions |
|---|
Intervention receipts
Risk decisions, predicted impact, resulting usage, and developer outcomes.
| Time | Tool / project | Decision | Risk change | Result |
|---|
Spend and improvement signals
Ranked by money, context pressure, and what to change next.
Read-only local scan · no model calls unless configured · no source or prompt content in summaries · nothing leaves this machine unless you enable a connected workflow.
Settings
Updates, AI Assist, trust, and install steps.
Companion
Ambient Watch starts with the dashboard. Standalone mode: aiwatcher watch --notify --overlay --interval 60.
Updates
Source checkouts can check GitHub here. Package installs update through their installer.
AI Assist
Optional local or cloud model help for Fresh Start and Plan. AIWatcher works without this.
Trust boundary
What AIWatcher can see and when data can leave the machine.
Surface coverage
Detected tools, hook confidence, and what AIWatcher can honestly measure.
Setup steps
Install and verify the local surfaces you want AIWatcher to protect.