atlasbound
v0.24.2
Published
AtlasBound , AI agents, bound to your codebase. Local-first closed-world code intelligence for AI coding agents (extract -> gate -> serve). Zero runtime dependencies.
Maintainers
Readme
AtlasBound
Product repo for AtlasBound: a local-first, closed-world code-intelligence platform that locks AI coding agents to a codebase's real architecture (extract -> gate -> serve), with a thin SaaS control plane (signup + license + redacted telemetry). Raw source code never leaves the developer's machine (an opt-in debug/tester tier can send a scrubbed dev-trace , PRIVACY.md §2b); DENY authority comes only from deterministic, parser-derived facts.
Quickstart
npm install -g atlasbound # or: npx atlasbound <cmd>
atlasbound setup # detect stack -> scaffold .atlasbound -> build -> wire hooks -> L1 verdict
atlasbound mcp wire # expose the mcp__atlasbound__* tools to your AI client (restart the session after)
atlasbound intake-init # install the intake + resume/handoff prompts into .claude/
atlasbound onboard # the live-measured checklist , do exactly NEXT, call again, repeat to 8/8
atlasbound update # refresh the snapshot after code changes (honest no-op when already fresh)Every L1-local feature works fully WITHOUT a cloud key; a key only adds licensing and redacted, allowlisted telemetry. Raw source code never leaves your machine; only redacted telemetry does (and, under an opt-in debug/tester tier, a scrubbed dev-trace , PRIVACY.md §2b).
Three pillars
- code intelligence: ground truth from your repo (closed-world extract -> gate -> serve).
- intake: mandatory pre-work interview + scope lock + confidence loop.
- handoff: bounded, parent-linked development continuity across /clear (decisions + checkpoint chain + critical artifacts + atomically rehydrated pack).
Setup/scaffolding is plumbing (an MCP setup tool), not a pillar.
