agent-mesh-core

agent-mesh-core is a personal AgentMesh reference implementation for coordinating trusted local AI agents across machines. A Mac mini owns the live mesh directory on local disk; Claude Code, Codex, and Ollama-side tooling talk to it over a private Tailscale network through MCP and HTTP services.

This is experimental personal infrastructure, not a general-purpose multi-tenant coordination service. The trust boundary is your private tailnet. Callers are trusted; the app layer does not authenticate users, authorize agents, or prevent one trusted caller from passing another agent_id.

What It Does

• Maintains per-agent inboxes, state files, and directory-based locks under a local mesh root.

• Writes JSON atomically with local-disk os.replace semantics.

• Uses claim-then-acknowledge inbox processing so dropped network clients do not silently lose claimed messages.

• Exposes the same dispatch layer through agent-mesh-mcp-server for Claude Code/Codex and agent-mesh-http-server for tools that cannot speak MCP.

Background and design rationale:

• Problem statement

• Implementation plan

• Operations guide

• Roadmap

Related MCP server: agent-bus-mcp

Trust Model

This project assumes one human operator, trusted machines on a private Tailscale network, and trusted local tools that are allowed to call the MCP/HTTP services.

This project does not provide multi-user authorization, per-client authentication, agent ID anti-spoofing, or a security boundary between tools already allowed to reach the service. If you need those properties, treat this repo as a design reference rather than deployable software.

Architecture

The live mesh root is local to the Mac mini:

AgentMesh/
├── config/
│   └── local_rules.json
├── agents/
│   ├── agent_mac_mini/
│   │   ├── inbox/
│   │   └── state.json
│   ├── agent_mbp/
│   │   ├── inbox/
│   │   └── state.json
│   └── agent_ollama_local/
│       ├── inbox/
│       └── state.json
└── locks/

Other machines never mount or write this directory. They call services running on the Mac mini:

• MCP: http://<mac-mini-tailscale-host>:8000/mcp

• HTTP wrapper: http://<mac-mini-tailscale-host>:8001

New Client Setup

Use this when the Mac mini services are already deployed and you want a new Mac, Linux, or Windows machine to talk to the mesh.

1. Install and log into Tailscale on the new machine.

2. Confirm the machine can reach the Mac mini Tailscale IP or MagicDNS name:

   tailscale status

3. Register the MCP server with Claude Code:

   claude mcp add --transport http agent-mesh http://<mac-mini-tailscale-host>:8000/mcp
   claude mcp list

4. Register the MCP server with Codex:

   codex mcp add agent-mesh --url http://<mac-mini-tailscale-host>:8000/mcp

Use the Mac mini's Tailscale IP if startup-time DNS resolution is a concern. MagicDNS is easier to read, but a raw tailnet IP avoids depending on DNS during launchd startup.

Per-OS Notes

The MCP registration commands are the same across operating systems. The platform-specific part is getting Tailscale installed and authenticated.

• macOS: install Tailscale from the App Store, direct download, or package manager; log into the same tailnet as the Mac mini.

• Linux: install Tailscale using the distribution instructions from Tailscale, start the service, and run tailscale up.

• Windows: install Tailscale for Windows, log into the same tailnet, and run the Codex/Claude commands from a shell that has those CLIs on PATH.

No SMB mount is needed on any client OS.

Mac Mini Setup

Use this when setting up the host role from scratch or doing local development.

1. Clone the repo and install dependencies:

   git clone <repo-url>
   cd agent-mesh-core
   uv sync

2. Run tests and lint:

   UV_CACHE_DIR=.uv-cache uv run pytest -q
   UV_CACHE_DIR=.uv-cache uv run ruff check .

3. Bootstrap a mesh root:

   MESH_ROOT=/Users/Shared/AgentMesh \
   AGENT_IDS=agent_mac_mini,agent_mbp,agent_ollama_local \
   ./deploy.sh

4. Render and load the MCP and HTTP launchd services.

   See docs/OPERATIONS.md for the exact commands, logging paths, and validation steps.

Ollama Integration

Ollama does not speak MCP and does not provide a built-in hook that makes arbitrary coordination calls before or after inference. The deployed HTTP wrapper exists for Ollama-side glue code to call explicitly.

Expected integration shape:

1. Keep agent-mesh-http-server running on the Mac mini.

2. Write or use an Ollama-side script/proxy that calls HTTP routes such as POST /send_message, POST /claim_inbox_messages, POST /acknowledge_claims, POST /read_local_rules, and POST /health_check.

3. Have that script identify itself with the provisioned agent ID agent_ollama_local.

Example health check:

curl -sS -X POST \
  http://<mac-mini-tailscale-host>:8001/health_check \
  -H "Content-Type: application/json" \
  -d '{}'

Example send:

curl -sS -X POST \
  http://<mac-mini-tailscale-host>:8001/send_message \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "agent_ollama_local",
    "target_agent_id": "agent_mbp",
    "message_type": "task.note",
    "body": {"text": "hello from ollama-side tooling"}
  }'

A reference poller that closes this loop end-to-end — claiming inbox messages, routing them to a local Ollama model, and replying to the sender — lives at examples/ollama/mesh_poller.py, with the task-routing rationale (what's a good fit for Ollama vs. what should stay on Claude Code/Codex) in examples/ollama/README.md. It's example code, not a deployed service — copy it into your own always-on process (launchd/systemd) if you want it running continuously.

Development

Useful commands:

UV_CACHE_DIR=.uv-cache uv run pytest -q
UV_CACHE_DIR=.uv-cache uv run ruff check .
UV_CACHE_DIR=.uv-cache uv run agent-mesh-bootstrap --mesh-root /tmp/AgentMesh --agent-ids agent_mac_mini,agent_mbp,agent_ollama_local
UV_CACHE_DIR=.uv-cache uv run agent-mesh-recover-processing --mesh-root /tmp/AgentMesh --agent-id agent_mbp --dry-run

Runtime paths such as agents/, locks/, and live config/*.json files are ignored by git. Keep the live mesh as data-only; source code and templates belong in this repository.

Status & Support

This is experimental personal infrastructure, maintained in spare time for a single operator's own machines — not a supported product with an SLA. Issues and PRs are welcome, but there's no commitment on response time, and feature requests outside the single-operator/private-tailnet scope described in Trust Model are likely to be declined rather than accepted. See SECURITY.md for the trust boundary and how to report a real security concern.

License

MIT