agentmesh-mcp 0.1.0

Multi-agent framework with persistent expert sessions and context isolation — MCP server for Claude Code

agentmesh

Multi-agent framework with persistent expert sessions and context isolation — runs as an MCP server for Claude Code.

The problem

When investigating complex bugs or features that span multiple domains (permissions, database, frontend...), a single Claude Code session accumulates too much context. It gets slow, expensive, and loses focus.

Why existing approaches don't fully solve it

Documentation files (SKILL.md, README, notes) give the agent orientation — but they load into your context, not a separate one. Every file you add to help Claude understand a domain is context your session consumes. And they're static: they don't update as Claude investigates.

Memory files have the same problem. They're summaries you write manually after the fact. They help with orientation but they don't capture the actual investigation — the grep results, the file reads, the chain of reasoning. Next session you start from the summary, not from where the investigation left off.

Neither approach isolates context. When Claude investigates a permissions issue to answer a question about pipelines, all that permissions code ends up in the pipeline session. Context bleeds.

The real problem is: there's no way to ask "what does the permissions expert know?" without loading all of that knowledge into the current session.

How agentmesh solves it

You define expert agents — each specialized in one domain. When Claude Code needs to know something outside its current focus, it calls query_expert. The expert:

1. Loads its persistent session (accumulated knowledge from prior investigations)
2. Investigates using tools (read files, grep, find) — in its own isolated context
3. Returns only the answer to Claude Code — never the investigation steps

Claude Code's context stays clean. The expert's context grows richer over time.

Claude Code (main agent):
  "Why does checkout fail for guest users?"
  → calls query_expert("auth", "can guest users reach the checkout flow?")

agentmesh expert (auth):
  [loads prior session: 6 messages of accumulated knowledge]
  [reads GuestSessionMiddleware.php]
  [greps for guest user checks in checkout]
  → returns: "Guest users are blocked by RequiresAccount middleware mounted on /checkout/*"

Claude Code receives only: "Guest users are blocked by RequiresAccount middleware mounted on /checkout/*"

What makes it different

Property	SKILL.md / docs	Memory files	AutoGen / CrewAI	agentmesh
Doesn't load into your context	no	no	no	yes
Captures live investigation results	no	manual	no	yes (automatic)
Persists across sessions	yes (static)	yes (static)	no	yes (grows with use)
Context isolation between domains	no	no	no	yes (core design)
Works with Claude Code as-is	yes	yes	no	yes (MCP)

Installation

pip install agentmesh

Quickstart

# 1. In your project root
cd my-project
agentmesh init
# → creates agentmesh.yml in the current directory

# 2. Edit agentmesh.yml (created in my-project/agentmesh.yml) to define your experts
# 3. Register with Claude Code:
claude mcp add --transport stdio agentmesh -- agentmesh serve --config /path/to/my-project/agentmesh.yml

# 4. Verify Claude Code picked it up
claude mcp list
# → agentmesh   stdio   /path/to/agentmesh serve ...

# 5. Test an expert directly (no Claude Code needed)
agentmesh ask orders "How does the order cancellation flow work?"

Configuration (agentmesh.yml)

model: claude-sonnet-4-6
session_store: .agentmesh/sessions   # where sessions persist (gitignored)
working_dir: .                       # root for file reads and shell commands

experts:
  orders:
    description: "Expert in order lifecycle, cancellations, refunds, and fulfilment logic"
    seed: ./docs/orders.md           # optional: pre-populate with context
    tools:
      - read_file
      - bash_readonly               # grep, find, cat, git log, ls

  payments:
    description: "Expert in payment processing, providers, retries, and webhooks"
    tools:
      - read_file
      - bash_readonly

  auth:
    description: "Expert in authentication, sessions, guest users, and access control"
    tools:
      - read_file
      - bash_readonly

  frontend:
    description: "Expert in React components, state management, and UI patterns"
    tools:
      - read_file

CLI reference

agentmesh init                         # create agentmesh.yml
agentmesh serve                        # start MCP server (used by Claude Code)
agentmesh ask orders "How does order cancellation work?"   # query an expert directly
agentmesh sessions list                # view all sessions with stats
agentmesh sessions clear payments      # reset a domain's accumulated knowledge

How sessions work

Each expert's conversation history is stored in .agentmesh/sessions/<domain>.json. Sessions persist across Claude Code invocations — the expert gets smarter over time without repeating investigations.

Add .agentmesh/ to your .gitignore (sessions contain investigation artifacts, not source code).

Built-in tools for experts

Tool	What it does	Restrictions
read_file	Read any file relative to working_dir	Read-only
bash_readonly	Run shell commands	Allowlist: grep, find, cat, ls, git log, git show, git diff, git blame, wc, head, tail

Examples

See examples/ for ready-to-use configurations:

• basic/ — single general-purpose expert
• codebase-investigation/ — permissions, database, frontend, API experts

License

MIT