Hermes Agent + Graphify: Codebase Knowledge Graphs

What Graphify Is

Graphify is MIT-licensed and builds a graph from a repository — code, docs, papers and diagrams. It uses Tree-sitter for static analysis, Leiden clustering to find community structure, and an LLM pass to attach semantic labels. Nodes are functions, files and concepts; edges are calls, imports and semantic relationships.

It launched in April 2026 and crossed 22,000 GitHub stars in under ten days, with native integrations across a range of AI coding platforms. Its headline number is the token saving: roughly 1.7k tokens per query against about 123k for a naive read-the-repo approach, a 71x reduction.

Why Hermes Specifically Benefits

Hermes Agent is strong at long-running autonomous work — it holds a persistent workspace, runs background reviews, and can grind through multi-step tasks without being re-prompted. That amplifies both sides of the retrieval problem.

An agent that works for an hour on a repository will re-read the same files many times if the only retrieval it has is grep-and-read. Every one of those reads is paid for, and each one crowds out context that mattered. A persistent graph inverts it: the agent asks a question, gets back the few relevant chunks, and spends its context on reasoning instead of on file contents it already saw twice.

Three concrete effects:

• Cheaper long sessions — the saving compounds across a multi-hour autonomous run rather than applying once
• Less context thrash — the agent stops evicting useful state to make room for source files
• Code stays local — graph construction runs on your machine, and only the query result travels to the model

Step 1: Build the Graph

From your repository root:

pip install graphify
graphify build .

Graphify walks the tree, runs Tree-sitter over code files, clusters the result, and writes an artefact under .graphify/. The build happens locally — no source is uploaded anywhere.

Step 2: Serve It Over MCP

graphify mcp serve --port 7812

This exposes the graph query tool over the Model Context Protocol, which is the same protocol Hermes uses for external tools.

Step 3: Register the Server With Hermes

Add the endpoint to your Hermes MCP configuration so the agent discovers the tool at startup:

{
  "mcpServers": {
    "graphify": {
      "url": "http://127.0.0.1:7812",
      "transport": "http"
    }
  }
}

Restart the agent, and the graph query tool shows up in its tool list. General MCP mechanics for Hermes — where config lives, how to debug a server that will not start — are in the Hermes MCP guide.

Reachability for a Hosted Hermes

A graph served on 127.0.0.1 of your laptop is not visible to a hosted container. If you run Hermes on OpenClaw Launch, you have the usual three options:

• Tailscale — put the agent and the machine serving the graph on one tailnet
• Run Graphify beside the agent — build the graph on the same host, which also keeps the repository there
• Publish it over authenticated HTTPS — workable, but you are now exposing a service that has read your entire codebase, so authenticate it properly

For a private repository, the second option is usually the honest one: if the code cannot leave your infrastructure, neither should the graph built from it.

Keeping the Graph Fresh

The graph is a snapshot. On an active repository it drifts, and a stale graph produces confidently wrong file references — the worst failure mode, because it looks like a correct answer. Rebuild on a schedule:

# nightly rebuild at 04:00
0 4 * * * cd /srv/repo && git pull --quiet && graphify build . >/dev/null 2>&1

Or have the agent do it itself with a scheduled job, which has the advantage that a failed rebuild can message you rather than failing silently.

Hermes or OpenClaw for This?

Both frameworks speak MCP, so Graphify works with either. Choose based on the workload rather than the integration: Hermes for long autonomous runs over a codebase, OpenClaw for chat-first help across many channels. If you want the OpenClaw version of this setup, see OpenClaw + Graphify; for the broader difference, see OpenClaw Launch vs Hermes Agent.

Troubleshooting

The agent never calls the graph tool

Usually tool discovery, not reasoning. Confirm the MCP server is listed in the agent's available tools after restart; if it is absent, the server did not start and the error is in the agent log rather than in chat.

Answers cite files that no longer exist

A stale graph. Rebuild it, then add the scheduled rebuild above so it does not recur.

The build is slow on a large monorepo

Expected on the first pass — Tree-sitter parsing and clustering are proportional to repository size. Subsequent builds are cheaper, and it is worth scoping the build to the directories the agent actually works in.

What's Next?

• OpenClaw + Graphify — the same integration on the OpenClaw side
• Hermes MCP guide
• AI agent cron jobs
• Managed Hermes hosting
• OpenClaw Launch vs Hermes Agent