Editor's note
ForestX opened its planner to outside agents.
Model Context Protocol is a plain idea with a large consequence: a common way for software to describe what it can do, so any AI assistant can call those functions without a separate integration built for each pairing. BTB first covered the protocol in #012, when it was still a curiosity in mapping tools.
The everyday version looks like this. An operator wants a delivery date, then wants to know what to ship instead when that date does not hold. Today that is two lookups inside the planning system. Over the protocol it is one question to whichever AI assistant is already open, answered without opening the planner at all. The software keeps doing the work while something else owns the screen.
Axel
PS. I am building one of these for SkogsAI. More on that shortly.
ForestX makes its sawmill planner callable by outside agents
The build is an interface layer: Optinex functions and data exposed as individually callable skills over Model Context Protocol, the open standard that lets an assistant invoke software directly. An agent can chain them, so a delivery-date lookup and a substitute-product query resolve in one request instead of two screens. ForestX says any assistant speaking the protocol qualifies, SAP's Joule included.
A stratospheric balloon reads fire risk over Ontario before anything burns
The ÉTS prototype Borealis flew on a Canadian Space Agency balloon out of Timmins during the ICARUS 2026 campaign in August, carrying cameras and an AI payload that reads the forest below and flags where fire risk is building before anything ignites. Student engineers and researchers at the LASSENA lab in Montréal built it. That places it upstream of every detection system this newsletter usually covers.
A segmentation model separates Christmas tree plantations from the forest around them
Feeding the model the vegetation it keeps mistaking for plantation conifers is what made this work: hard negative mining lifted the precision-recall score from 0.204 to 0.913. A DeepLabV3 segmentation network with a ResNet-34 encoder then mapped 87,309 hectares of the French Morvan, scoring 0.846 on 2020 imagery and holding 0.817 on 2023 without retraining. Cross-year transfer is what an operational monitoring product gets judged on.
Terrain-aware competition indices lift per-tree volume accuracy in Chinese fir
Backpack lidar returns from #060, fused this time with drone lidar and aimed at volume rather than area. The group used a terrain model to decide which neighbouring trees genuinely compete for light and space, instead of which merely sit close on a flat map. Across 12 Chinese fir plots the share of variation the model explains rose from 0.74 to 0.85, and from 0.81 to 0.90 in dense stands. A tree's own dimensions still dominate; slope-aware competition adds at the margin.
Foster + Partners joins a robot swarm that builds timber floors from the ground up
Foster + Partners has put its name to SWIFT-BUILD, a project the European Innovation Council cleared on 1 April under its Pathfinder call for autonomous robot collectives in construction. University of Bristol coordinates, 36 months, budget put at 4 million euros in press material. The method inverts the usual sequence: each floor is assembled at ground level and lifted into place, with drones overhead flagging deviations live. Timber construction rather than forestry.