pandō — send less code to your AI provider
68%less code exposure overall
across three VS Code slices
Change the name.
Keep the identity.
Rename connected references in one request. Every committed change is saved automatically.
export function getUser(id: string) {
return users.get(id);
}
const profile = getUser(userId);browse-code → symbolId: "73001"Find the symbol once. Its identity stays the same when its name changes.
What is pandō?
LESS CODE EXPOSED
A “code firewall”
for your AI agent.
“Add Azure and Google Cloud support to Northstar.”
lines of your code
never sent to OpenAI in this example.
Illustrative line counts for a fictional app, not a measured Northstar result. The VS Code benchmark measured 68% less exposure in bytes; actual savings vary.
The same change. Less code reaching the model provider. That’s the “code firewall” metaphor: minimizing exposure, not eliminating it.
How does it work?
YOUR CODE, CAPTURED ONCE
Your program is a database.
Your tools still see files.
- 01
Capture your project
pandō captures your project in a local program store while leaving your original checkout untouched.
- 02
Work on the graph
Your agent navigates code as a graph and makes coordinated changes, with each committed step saved automatically.
- 03
Keep your existing tools
pandō presents your program as ordinary files, so classic tools, builds, and tests can keep working.
The original checkout and pandō workspace can diverge. They are not automatically synchronized. Workspace guide ↗
What can my agent do?
MORE ROOM TO THINK
Focus on the problem.
Carry the change through.
You set the direction. Your agent can spend its attention on design and behavior, with tools that handle connected edits and save each committed step.
- 01
Follow the connections
Navigate code as a graph: ask for callers, references, and dependencies directly. Read the relevant signature or body when you need it. Spend less effort reconstructing relationships from text.
- 02
Carry the change through
Rename a symbol and its bound references across the project. Change parameters and their calls together. Move or wrap existing code without retyping it.
- 03
Keep track as code changes
Your agent can return to the same construct after a rename or move. It doesn’t have to rely on a line number that has shifted.
Operations report what applied or refused. Tests still check behavior. Try a rename example ↗
Can I go back?
EVERY STEP SAVED
Go back.
Try another path.
Your agent never has to remember to save a checkpoint. Each committed structural change automatically creates a semantic commit, preserving the program and its relationships. No extra commit step or bookkeeping.
history({ "select": "17" })Want to try a different path? Select any earlier snapshot with one call, then keep working from there. No undo sequence or reverse patches needed. The path you left stays in history, ready to revisit.
Builds and behavioral tests still matter. Read the history contract ↗
Who sees your code?
KEEP MORE CODE PRIVATE
Give your agent context.
Expose less of your code.
less code exposure overall
across three VS Code benchmark slices
Your agent can understand connections and make structural changes without routinely reading and rewriting entire function bodies. Focused reads let it request the code it needs for the task.
September 27, 2026 report, versus a text-tool agent, with pandō memory disabled. Across three slices: 812,294 → 259,970 bytes. Exposure counts code in tool outputs plus agent-written code retained in the repository; only conversation resends are deduplicated. Reasoning and harness content are excluded. These are task-specific, model-labelled measurements; full code reads still disclose source. This is a code-exposure result, not a speed claim. Work-time comparisons exclude pandō’s roughly 15-minute, one-time initial indexing step; a two-hour work session still pays that setup cost only once, plus the time taken by the work and each operation. Download the full benchmark report (.xlsx) ↗
TS / JS One rename. Connected references. An automatic checkpoint.
/ˈpan.doː/ · “I spread” · from Latin. Named after the world’s largest organism — a quaking aspen grove in Utah, 47,000 trees connected by a single root system. One connected program, many rendered views.