Today we are shipping Bootspring v3.0.0 — the first major version bump since 2.7.0 and the largest architectural leap in the project's history. This release introduces swarm intelligence: the ability to coordinate multiple AI agents working together on your codebase with real-time conflict detection, shared memory, and goal-driven planning.
Why a Major Version
v3.0 represents a fundamental shift in how Bootspring operates. Previous versions treated each agent invocation as an isolated request. v3.0 introduces persistent, coordinated agent groups that share context, vote on decisions, and self-heal when agents fail. This is not a point release — it changes the mental model for how you interact with the platform.
Headline Features
Swarm Intelligence
Launch a coordinated group of agents that work together on complex tasks. Five orchestration topologies let you match the coordination pattern to your problem:
- Hierarchical — A lead agent delegates subtasks to specialists and aggregates results.
- Mesh — All agents communicate freely, ideal for brainstorming and exploration.
- Ring — Agents pass work sequentially, refining output at each stage.
- Star — A central coordinator dispatches to and collects from worker agents.
- Adaptive — The topology shifts automatically based on task characteristics.
Swarms include consensus voting (majority, supermajority, unanimous), heartbeat monitoring, and automatic self-healing when an agent becomes unresponsive.
Via the CLI:
Multi-Model Deep Integration
Assign specific AI models to individual agents. Run Claude Opus for architecture decisions, Sonnet for implementation, Haiku for tests and linting. Connect local models via Ollama for air-gapped environments.
Per-agent cost tracking gives you granular visibility into spend:
Automatic Memory Recall
Every agent turn now triggers a passive vector search against the project's memory store. Relevant memories — past decisions, architectural patterns, known issues — are injected into the agent's context automatically. No explicit tool call needed.
This means agents remember what was decided three weeks ago without you restating it. The HNSW-indexed vector store provides sub-5ms retrieval even with 10,000+ stored memories.
File-Conflict Detection
When multiple swarm agents are active, Bootspring tracks file reads and writes in real time. If Agent B writes to a file that Agent A previously read (and may have stale assumptions about), both agents are notified immediately.
Structure-Aware Search
Code search now returns the enclosing function or class context, not just matching lines. Supported languages: TypeScript, JavaScript, Python, Rust, Go.
Federation
For teams running multiple Bootspring instances, federation enables cross-node swarm coordination with JWT-authenticated communication, task migration, and distributed consensus.
By the Numbers
| Metric | v2.7.0 | v3.0.0 |
|---|---|---|
| Tests passing | 7,840 | 9,220 |
| Test files | 298 | 344 |
| CLI commands | 76 | 81 |
| MCP tools | 34 | 40 |
| Expert agents | 37 | 43 |
| API endpoints | 180+ | 240+ |
| Package size (packed) | 980 KB | 1.1 MB |
The thin-client architecture is preserved. All swarm orchestration, vector memory, and model routing runs server-side. The npm package ships 17 files — CLI entrypoint, MCP server, and core utilities only.
Upgrade Path
v3.0 is backward-compatible for existing CLI and MCP workflows. Your bootspring build, bootspring seed, and bootspring agent commands work exactly as before. Swarm features are additive — you opt in by creating a swarm.
What's Next
We are actively building toward v3.1:
- Real-time collaboration — WebSocket-based live swarm dashboard with shared cursors
- Visual workflow builder — Drag-and-drop DAG builder for GOAP plans
- Cost budgets — Per-swarm spend caps with automatic model downgrade
- Extended language support — Structure-aware search for Java, C#, Ruby, PHP
Get Started
Bootspring v3.0.0 is available now on npm. Questions? Join our Discord or check the documentation.