meals rescued↗ 18.4%
2,418
this month
rescue control room v1.1 / 13 webmcp tools
Move good food
before it's too late.
FoodLoop turns surplus meals into coordinated community handoffs. Let an agent find the best fit, then keep the human in control of every dispatch.
agent-ready
↗
match
route
handoff
WebMCP layer
Structured tools
for human judgment detecting browser support
for human judgment detecting browser support
01Explainable matchingdiet · deadline · distance
02Constraint-aware routeprecedence · capacity · slack
03Two human gatesdispatch · final receipt
04Inspectable trailtool calls · decisions · impact
tonight's supply
346
servings available live
co₂ avoided↗ on pace
1,420kg
equivalent to 3,520 mi
needs attention
2urgent
6 open requests
judge proof / safety lab
Make the reasoning inspectable.
One shared state machine for agent calls, refusals, counterfactuals, and the human decisions that close the loop.
fresh-browser walkthrough
See the agent reason — and stop.
Tool calls are checked against this capability boundary.
Suggested fresh-browser prompt
Read context → chain a safe plan → try the protected action as Planner.
Role-scoped capabilitiesactive role highlighted
Agent tracecalls + refusals
adversarial safety lab
Refuse unsafe work with evidence.
Every scenario is evaluated in an isolated sandbox. The refusal is visible; the live route stays unchanged.
decision intelligence
Ask what changes.
Match explanationS-104 → R-201
Route stress testWhat if departure slips?
impact ledger
Measure the operating difference.
Illustrative comparison for the Springfield simulation — not field research or a claim about real-world outcomes.
coordination baseline42 minsimulated manual workflow
agent-assisted—scenario estimate
time saved—vs. baseline
servings rescued0verified handoffs
unsafe plans rejected0sandbox refusals
handoff completion0%verified / staged
route efficiency—miles / 100 servings
vehicle utilization—load / capacity
Run the scenario to compare the assisted workflow with the simulated baseline.
external-data benchmark
Does the route hold up outside Springfield?
A reproducible adapter benchmark on the public FoodGo food-rescue dataset. The product demo remains fictional; this table tests the same deadline, capacity, and route-quality instincts against published records.
Loading the public route benchmark…Reading the checked-in FoodGo dataset result.
Source: USDOT / CMU FoodGo Test Dataset · CC BY 4.0 · reproduce with
npm run benchmark
critical path
The agent proposes. You decide.
- 01Intakecapture supply
- 02Matchfit needs
- 03Dispatchdraft route
- 04Handoffclose the loop
natural-language intake
Tell the control room what changed.
Try a scenario
incoming supply
Surplus lots
community signal
Recipient requests
operational picture
Tonight's rescue network
NORTH END
RIVERSIDE
EAST MARKET
DOWNTOWN
No active route yetRun the rescue scenario to plot stops.
decision queue