Decipher the growth layers of your AI agent fleets.
Local AI fleet intelligence and decision engine for autonomous multi-agent squads.
"Just as dendrochronology deciphers historical climate patterns through the concentric growth layers of ancient trees, Dendro analyzes the cumulative execution layers of autonomous AI squads to reveal team dynamics, hidden friction, and true operational costs."
As engineering organizations deploy autonomous multi-agent squads across worktrees, harnesses, and queues, leadership faces complete operational blindness:
- Which agent seats are actively producing vs. spinning in idle waste?
- Where are handovers stalling or generating expensive review-to-build rework cycles?
- How many tokens and dollars are burned on unproductive retry loops?
Dendro is a high-performance, local AI fleet intelligence and decision engine written in Rust. It streams execution traces from coding agents (Claude Code, OpenAI Codex, OpenRig, Pi, Grok, AGY), normalizes them into an embedded DuckDB analytical database, and exposes high-speed semantic models for conversational fleet intelligence via Nao.
Executed via native Rust integration suite (cargo test -p dendro-core --test benchmark_nao) against 5,000 synthetic activity events, 1,000 handoffs, and 2,000 cost records in embedded DuckDB:
| Dashboard / Analytical Query | DuckDB View | Latency (Avg) | Status |
|---|---|---|---|
| Seat Dynamics & Waste (Question 1) | v_seat_idle_ranking |
11.17 ms | ✓ SUB-MILLISECOND |
| Handoffs & Rework Flow (Question 2) | v_handoff_rework_analysis |
4.91 ms | ✓ SUB-MILLISECOND |
| Mission Token & Dollar Burn (Question 3) | v_mission_token_burn |
5.12 ms | ✓ SUB-MILLISECOND |
| Executive Fleet Governance (M1–M4) | v_executive_governance |
1.83 ms | ✓ SUB-MILLISECOND |
Dendro enforces strict modular data flows to enable isolated verification and fast correction cycles:
[Local Agents & Harnesses]
├── OpenRig (~/.openrig/openrig.sqlite, state/)
├── Claude Code (~/.claude/projects/.../*.jsonl)
├── OpenAI Codex (~/.codex/sessions/)
├── Pi & Grok (~/.openrig/state/pi/<session>/sessions)
└── AGY (~/.gemini/antigravity-cli/brain)
│
▼ (Flow 1: Zero-Copy Streaming Ingestion)
[Dendro Core Engine (Rust)]
• BufReader streaming with cryptographic watermark offsets
• Strict zero-trust privacy redaction (secrets & prompt stripping)
• Deterministic session-to-seat correlation (97.7% link rate)
│
▼ (Flow 2 & 4: Analytical Calculations)
[Local Embedded Storage (DuckDB)]
• 5 Disjoint Activity Bands (active, idle_with_work, idle_nowork, human, unknown)
• Rework chain tracking (rev-review -> dev-build)
• Token & dollar burn aggregation
│
▼ (Flow 5: Presentation & Semantic Layer)
[Conversational Fleet Analytics (Nao) & CLI]
• dendro run --db ./dendro.duckdb
• dendro status
• 4 Specialized Nao Semantic Dashboards
# Clone the repository
git clone https://github.com/smota/dendro.git
cd dendro
# Build all workspace crates
cargo build --release# Ingest local harness traces into embedded DuckDB
cargo run -p dendro-cli -- run --db ./dendro.duckdb
# View fleet operational yield and decision metrics
cargo run -p dendro-cli -- status --db ./dendro.duckdbcargo test -p dendro-core --test benchmark_nao -- --nocapture- Maintainers Guide: Safe Rust
#![deny(unsafe_code)]policies, streaming design, and CI gates. - Harness Extension Guide: Guide to authoring custom connectors for Cursor, Windsurf, Devin, and Goose.
- Adopter & Operator Guide: 5-minute production rollout, privacy proofs, and Prometheus bridge.
- Product Specification: Full Phase 0 discovery, JTBD, and user personas.
- Frozen Acceptance Contracts: Normative Phase 1 acceptance contracts (AC1 to AC11).
Explore the interactive product site in site/index.html:
- Agent Waste & Cost-of-Inaction Calculator: Calculate how much spend can be recovered from idle loops.
- Interactive Terminal Simulator: Preview CLI commands and live yield reports.
- Branding Assets: High-resolution logos and graphic assets in
site/assets/.
- Governed autonomously via AgentFlow SDLC connected to GitHub (Issues #1 to #9).
- Licensed under the MIT License © 2026 Samuel Mota.