---

✨ What is it

AGNOHive is a swarm of specialized agents (Researcher, Planner, Coder, Reviewer, …) that connect to your codebase through MCP and get real engineering work done — reading files, planning changes, writing diffs, running commands — all orchestrated by a coordinator model, all running on hardware you control.

• 🔒 Nothing leaves your network — every model call is local (Ollama or vLLM), every file access is via your own MCP server
• 🧠 Grounded, not guessed — agents read the actual codebase before answering; a hive.md snapshot + LightRAG semantic index keep them from hallucinating structure
• ✅ Human-in-the-loop by default — every file write and every external-platform action is staged for your approval before it lands
• 🔌 Works with any project — point it at any repo via MCP, no project-specific setup required beyond the connection
• 🔀 Pluggable inference backend — Ollama or vLLM + LiteLLM, switchable with one env var, no code changes

🏗️ How it works

flowchart LR
    subgraph Client["💻 Client Machine"]
        HM["🐳 hive-mcp<br/>apply_diff · write_file<br/>run_shell · run_docker<br/>git_* · index_project<br/>scan_project_context → hive.md<br/>web_search / web_fetch"]
        PM["🔌 Project MCP<br/>get_file_content · find_files<br/>search_files · memory_search"]
    end

    subgraph ZGX["🖥️ Workstation (ZGX)"]
        COORD["🧠 Coordinator<br/>qwen3-coder:30b"]
        PIPE["ContextRouter → Researcher → Planner → Coder → Executor → Reviewer"]
        QD[("Qdrant<br/>vectors")]
        PG[("PostgreSQL/AGE<br/>graph")]
        SZ["SigNoz<br/>OTel traces"]
    end

    HM <-->|Tailscale| COORD
    PM <-->|Tailscale| COORD
    COORD --> PIPE
    PIPE --> QD
    PIPE --> PG
    COORD -.-> SZ

Two MCP connections per run: hive-mcp (primary — all reads/writes/shell/git/web) and your project MCP (supplementary — app-specific tools like memory_search). If hive-mcp is unreachable, agents fall back to project MCP automatically; if both are down, the run fails with a clear error.

1. Coordinator's first action is get_file_content('hive.md') — grounded context loaded on demand, not pre-injected (prevents models from answering without tool calls)
2. Failure context from past runs is injected into the coordinator's instructions
3. The coordinator routes each operation to the right MCP; member agents see only their scoped tool subset
4. After each run: successes → LightRAG (vector memory), failures → PostgreSQL (failure log), traces → SigNoz

📟 A quick look

$ hive
AGNOHive  project EkamApp  mode engineering  http://100.96.86.82:9001
  project:   http://100.87.159.1:9000/mcp   + 12ms
  hive-mcp:  http://100.87.159.1:9003/mcp   + 8ms
  resuming session a3f7c2d1  (last used this project)
  /new  /sessions  /history  /persist  /delete <id>  /diff  /cleanup  /mcp  /confirm  /reject  /exit