HermesAgent

HermesAgent is a self-improving, model-agnostic coding agent by Nous Research. It supports MCP natively — its client auto-discovers NestWeaver’s 40 tools at startup and registers them in the agent’s tool registry.

Setup

Option 1: Auto-setup (recommended)

nestweaver setup hermes

This writes a project-local .hermes/config.json with the NestWeaver MCP server entry, merged with any existing config.

Option 2: HermesAgent native config

Edit ~/.hermes/config.yaml and add NestWeaver under mcp_servers:

mcp_servers:
  nestweaver:
    command: 'nestweaver'
    args: ['mcp', '--db', '/path/to/your.lbug']
    env:
      NESTWEAVER_DB: '/path/to/your.lbug'
    timeout: 30
    enabled: true

The env block is important — HermesAgent only passes explicitly declared environment variables to stdio subprocesses, not the full shell environment.

Option 3: HermesAgent CLI

hermes mcp add nestweaver --command nestweaver --args '["mcp", "--db", "/path/to/your.lbug"]'

Reload without restarting

If you’re already in a HermesAgent session:

/reload-mcp

Tool naming

HermesAgent auto-discovers tools and registers them with the prefix pattern:

mcp_nestweaver_<tool_name>

For example:

NestWeaver tool	HermesAgent name
`brain_context`	`mcp_nestweaver_brain_context`
`brain_search`	`mcp_nestweaver_brain_search`
`read_symbols`	`mcp_nestweaver_read_symbols`
`blast_radius`	`mcp_nestweaver_blast_radius`
`flow_trace`	`mcp_nestweaver_flow_trace`

All 40 tools follow this pattern. The agent calls them by their prefixed name — no manual mapping needed.

Tool filtering

To limit which NestWeaver tools HermesAgent sees, use include/exclude lists:

mcp_servers:
  nestweaver:
    command: 'nestweaver'
    args: ['mcp', '--db', '/path/to/your.lbug']
    tools:
      include:
        - brain_context
        - brain_search
        - read_symbols
        - blast_radius
        - flow_trace
      exclude: []

When include is set, only listed tools are registered. An empty include (the default) means all tools.

Alternatively, use NestWeaver’s built-in lite mode:

mcp_servers:
  nestweaver:
    command: 'nestweaver'
    args: ['mcp', '--lite', '--db', '/path/to/your.lbug']

You can also disable resource and prompt utilities if you don’t need them:

tools:
  prompts: false
  resources: false

Skills integration

HermesAgent’s skills system learns from experience — after completing complex tasks, it creates reusable skill files in ~/.hermes/skills/. When the agent uses NestWeaver tools to solve a problem, it may generate skills that document effective NestWeaver query patterns.

For example, after a session where the agent uses brain_context with tag filtering to navigate a monorepo, it might create a skill documenting that workflow for future sessions.

Skills don’t wrap MCP tools directly — they’re procedural instructions the agent follows, and those instructions can reference any tool in its registry, including NestWeaver’s mcp_nestweaver_* tools.

What NestWeaver adds to HermesAgent

Without NestWeaver, HermesAgent navigates code using ripgrep and paginated file reads (capped at 100K characters per read). Each round-trip costs tokens — the LLM reasons about what to search, processes search results, reads files, then reasons again. With NestWeaver:

Single-call answers — brain_context returns the relevant symbols ranked by structural importance, within a token budget. No grep → read → re-grep loops.
Symbol-level precision — read_symbols returns a function’s source span (~400 tokens) instead of the entire file (~8,000 tokens for a 2000-line file).
Structural queries — “What calls this function?” answered from the graph in one call, not by reading 10+ files.
Persistent codebase memory — The graph persists across sessions. Session 100 starts as fast as session 1.
Reduced context compression — Less raw code means HermesAgent’s middle-message summarization triggers less often and loses less information.

Config reference

Full config options for the NestWeaver MCP server in HermesAgent:

mcp_servers:
  nestweaver:
    command: 'nestweaver' # executable
    args: ['mcp', '--db', '/path/to.lbug'] # CLI arguments
    env: # environment variables
      NESTWEAVER_DB: '/path/to.lbug'
    timeout: 30 # tool call timeout (seconds)
    connect_timeout: 10 # startup timeout (seconds)
    enabled: true # set false to disable without removing
    supports_parallel_tool_calls: false # concurrent tool execution
    tools:
      include: [] # tool whitelist (empty = all)
      exclude: [] # tool blacklist
      prompts: true # enable prompt utilities
      resources: true # enable resource utilities

Verify the connection

Start HermesAgent and check the tool registry:

hermes

At startup, you should see NestWeaver’s tools registered as mcp_nestweaver_*. Ask the agent to query the graph:

What are the main entry points in this codebase?

The agent should use mcp_nestweaver_brain_context rather than falling back to search_files and read_file.