Bubbly is a sovereign agentic coding environment: an IDE, a CLI, and a mission-control room where autonomous agents read your whole codebase, plan a change, write it across files, run the terminal and tests, debug failures, and review their own work. Every model runs locally on your own hardware via Ollama, so your source, prompts, and secrets never leave the building.
Read → Plan → Edit → Test → Review
code lives
Modern agentic dev tools can understand a codebase, write across files, run tests, and fix their own mistakes. Most share one problem: your proprietary source, your prompts, and your secrets are sent to someone else’s servers for any of it to work.
Bubbly delivers that autonomy as a local software factory. It is an agent runtime, a codebase index, and a tool sandbox, with a model in the middle, running entirely on infrastructure you control. For teams that cannot ship their source to a third party, that is the difference between adopting agentic engineering and being locked out of it.
reviewed code
Bubbly carries a task from intent to reviewed, tested code without skipping the steps a careful engineer would take.
Reads the whole codebase
Symbol, semantic, dependency, and call-graph retrieval build a working set for the task, not blind, file-by-file keyword matches.
Plans before it codes
Decomposes the task into a tracked plan with dependencies and a clear order of work before touching a single file.
Edits across files
AST-aware patches that rename symbols and update every call site automatically, never blind, destructive overwrites.
Runs the terminal and tests
Executes commands and test suites in an isolated sandbox. Work is not done until the tests pass.
Debugs to root cause
Traces a failure through the call graph and data flow to the actual cause instead of patching the symptom.
Reviews its own work
A critic agent checks for architecture drift, missing edge cases, and full spec coverage before anything is handed back.
Ollama
Every model Bubbly uses runs locally through Ollama on your own GPUs. There is no required cloud endpoint, no API key shipping your code to a vendor, and no telemetry phoning home. It runs fully air-gapped.
A model router picks the right local model for each step, so you get strong results without one giant model carrying every task. The result is coding autonomy with the data posture of an offline workstation: your weights, your hardware, your code.
four ways in
One session backbone, four ways to work, and you switch between them without losing state.
logged, reversible
Control
Approval gates before risky actions, bounded and explicit tool permissions, and a human in the loop by default.
Accountability
An append-only audit log where every action is attributable, with the full decision history preserved.
Safety
An isolated execution sandbox with snapshot and rollback, and network egress restricted to an allowlist.
Sovereignty
Runs inside your perimeter on your own models. Your code never leaves, and no external service is required.
Give your engineers an autonomous coding agent that understands the whole codebase and runs entirely on your own models, so the work moves fast and the source never leaves the building.