Fleet overwatch for multi-agent AI systems. A sleeping guardian on Cloudflare that gives your AI fleet self-awareness.
Version 0.1.0, no releases, created 2026-03-24 — approximately seven weeks old at appraisal. Active development with last commit 2026-05-10 (one day prior). Clearly alpha: roadmap items listed in README ("GitHub webhook integration," "fleet-bridge registry parsing," "biometric correlation layer") are unimplemented. The "built in one afternoon" origin story in README is candid about this. No issue tracker activity, no versioned releases, no changelog.
README is unusually strong for an alpha personal tool. Includes: problem statement with personal narrative, architecture diagram with ASCII art, full endpoint table split by auth tier, 5-minute deploy walkthrough, agent wiring examples with real curl commands, and cost breakdown ($5/month Workers plan, ~$0 Durable Objects at this scale). Gaps: no docs beyond README, no inline code comments visible, no API schema or OpenAPI spec, and the license field is empty despite README claiming MIT — a minor but real inconsistency.
TypeScript is appropriate for Cloudflare Workers. Dependency footprint is minimal and clean: agents ^0.0.50 (Cloudflare Agents SDK), wrangler ^4.0.0, and @cloudflare/workers-types — no unnecessary runtime bloat. Architecture is well-reasoned: Durable Object with embedded SQLite for five domain tables (repos, agents, agent_registry, alerts, checks) is correct for this use case. No tests are visible, no CI configuration mentioned, no linting setup declared. Code quality is inferred from design, not directly observable.
Single-developer personal project (NorthwoodsSentinel / "Rob," a security architect). Commit cadence is active — last commit one day before appraisal. Zero open issues, though this likely reflects low external usage rather than triage discipline. No PRs, no contributor graph. Maintenance continuity is plausible given author's stated daily-use motivation but is one person deep with no succession signal.
1 star (likely author), 0 forks, no downstream dependents discoverable, no releases, no package registry publication. Explicitly private ("private": true in package.json). Not seeking adoption — this is infrastructure the author built for himself and open-sourced. Adoption score reflects external uptake, not utility.
Overall: 2.4/5
Category: Multi-Agent Fleet Monitoring Known alternatives in vault: None. No existing vault entry addresses multi-agent coordination or fleet-level awareness. NorthwoodsSentinel--loam (Personal AI Memory, same author) is conceptually adjacent — both address agent context fragmentation — but loam operates at the individual session memory layer whereas brook operates at the fleet coordination layer. Differentiation: Brook is the only item in the vault targeting the multi-agent fragmentation problem from the coordination side: who is online, what did each agent build, when did they last check in, is anyone drifting. It is deliberately observer-only (no inference, no action), Cloudflare-native, and cost-optimized to near-zero. Alternatives in the broader ecosystem (LangSmith, AgentOps, Helicone) offer richer observability but require cloud accounts, cost more, and are not self-hosted. Brook trades depth for ownership and cost. Gap or crowd: Genuine gap. No vault entry currently provides inter-agent coordination or fleet-level awareness. The MCP integration on the roadmap would make this queryable by Claude directly, aligning with the vault's observed trend of MCP-first tooling (60% of vault projects).
Score: 4/5
Harvestable: Agent check-in/check-out REST pattern (session lifecycle hooks); drift detection logic (24-hour threshold, configurable); fleet registry schema (five-table SQLite design reusable for any multi-process tracking system); Cloudflare Durable Object + SQLite as cost-efficient persistent singleton pattern; hourly cron-on-Durable-Object scheduling approach for near-zero-cost background polling.
Integration path: Brook maps directly onto the coordination layer of a PAI system running multiple Claude Code instances. The /checkin and /publish endpoints can be wired into agent session start/end hooks. The /fleet and /agent/:name endpoints give any agent read access to the shared registry. The planned MCP endpoint would make brook natively queryable without curl wrappers. Deployment is genuinely five minutes given existing Cloudflare infrastructure.
Overlap with existing: NorthwoodsSentinel--loam addresses individual agent memory persistence; brook addresses fleet-level awareness — these are complementary layers, not duplicates. No other vault item overlaps with fleet coordination. Both are from the same author and likely designed to coexist in the Northwoods Sentinel Labs ecosystem.
Adoption cost: Trivial to deploy as infrastructure. Moderate to wire into existing agent workflows — each agent needs session-boundary curl calls, and the GitHub webhook requires a public endpoint and repo-level configuration. No code modification to existing agents beyond adding check-in/publish calls.
Same-author pair with NorthwoodsSentinel--loam: loam handles what an individual agent remembers; brook handles what the fleet knows. Together they address both halves of the context fragmentation problem in multi-agent PAI systems. The "built in one afternoon" origin is credible given the focused scope and minimal dependency footprint. The missing license file is a real gap for vault inclusion — README claims MIT but the repo has no LICENSE file; worth flagging to author. The biometric correlation roadmap item (Garmin/Oura integration) is an unusual signal — suggests the author intends brook to eventually correlate agent activity with operator physiological state, which would be a distinctive PAI differentiation if implemented. MCP endpoint on roadmap is the highest-leverage pending feature for vault integration purposes.