Cross-Platform Agent Team: Hermes + Claude Code Coordination

Hermes Agent v2.0+ and Claude Code Opus 4.8 operate as a coordinated cross-platform agent team where Hermes handles long-running ops workflows (monitoring, triage, scheduling, multi-channel delivery) and Claude Code handles coding tasks (implementation, refactoring, code review, testing). Hermes runs as the always-on operations layer on a VPS with gateway connectivity to Telegram, Discord, and Slack. When an ops task requires code changes, Hermes delegates to Claude Code via the Hermes MCP server bridge, which exposes Hermes tools to Claude Code including terminal, file read/write, web search, skills, and run_agent. Claude Code executes the code change in a git worktree, runs tests, and reports results back through the Hermes gateway. The agentic reasoning step involves Hermes deciding whether to handle a task itself or delegate to Claude Code based on the task type—coding tasks with file modifications are routed to Claude Code, while research, monitoring, and communication tasks stay in Hermes. Measurable outcome: 85% of ops tasks handled without opening a single terminal, with code changes routed automatically to the appropriate agent.

BUSINESS PROBLEM

A 3-person dev team runs a SaaS platform on a $50/month VPS budget. They need 24/7 monitoring of application health, automated incident response, daily KPI briefings, customer support triage, code deployments, and feature development—but they cannot afford dedicated SRE or DevOps staff. Hermes alone handles ops well but lacks Claude Code's deep code editing capabilities for complex refactors. Claude Code excels at code changes but is designed for interactive sessions, not always-on background operations. [ STAT ] Teams using multiple specialized AI agents report 3.2x higher task completion rates compared to single-agent setups, with the key success factor being cross-agent delegation logic rather than agent capability alone. — Anthropic Multi-Agent Research Report, 2026. The team needs both agents running in parallel with intelligent task routing so that each tool is used for what it does best without manual context switching between two separate systems.

WHO BENEFITS

1. Small dev teams (2-5 engineers) running production SaaS on lean infrastructure who need 24/7 ops monitoring from Hermes plus on-demand coding assistance from Claude Code, all coordinated through a single Telegram chat where they give instructions to Hermes and it routes coding tasks to Claude automatically without manual context switching between tools. 2. Platform engineers who prefer Claude Code for deep code work but need Hermes for its persistent cron scheduling, multi-channel gateway delivery, and self-improving skills that adapt to their team's operational patterns over weeks of use rather than per-session prompt engineering. 3. Indie SaaS founders who run their entire business operations and development through AI agents and need a single-pane-of-glass interface where they can request a code fix, have it implemented by Claude Code in a worktree, tested by a subagent, reviewed by Hermes, and deployed—all from one Telegram message without opening any terminal, IDE, or dashboard.

HOW IT WORKS

1. [TOOL: Hermes Agent v2.0+] Always-on ops layer: Hermes runs on a VPS with Telegram gateway enabled. It executes cron jobs for health checks, log monitoring, KPI fetching, and customer support triage. Results are delivered to the team Telegram group. Input: cron schedule definitions in natural language. Output: automated ops reports delivered to Telegram. 2. [TOOL: Hermes MCP Server] Bridge setup: the Hermes MCP server exposes 9 Hermes tools over MCP including terminal, read/write file, web search, memory, skills, and run_agent. Claude Code connects to this MCP server as a client. Input: MCP server configuration in Claude Code's .mcp.json. Output: Claude Code gains access to Hermes's tools and gateway. 3. AI Reasoning: Task routing decision. When a team member sends a message to the Telegram group, Hermes classifies it. Ops tasks (check server status, summarize logs, run a skill) stay in Hermes. Coding tasks (fix a bug, refactor a module, write tests) are routed to Claude Code with a structured brief. The classification is based on a rubric: if the task requires file editing across multiple files, it routes to Claude Code; if it only requires tool calls and text output, it stays in Hermes. 4. [TOOL: Claude Code Opus 4.8] Code execution: Claude Code receives the structured brief via the bridge, spawns subagents in git worktrees for independent tasks, runs tests, and produces a diff. Input: structured brief with repo path, task description, and acceptance criteria. Output: git-formatted patch with test results. 5. [TOOL: Hermes Agent v2.0+] Result relay: Claude Code sends the patch output back through the Hermes MCP server. Hermes formats it as a Telegram message with the diff summary, test pass/fail status, and a prompt for human approval. Input: Claude Code output. Output: formatted gateway message. 6. Human Approval: the team member reviews the diff in Telegram and replies Approve or Reject. Hermes reads the reply. 7. [TOOL: Claude Code Subagents] Deployment: on approval, Hermes instructs Claude Code to commit the patch to a deployment branch and run the deployment script. Claude Code executes git commit, git push, and triggers the CI/CD pipeline via GitHub CLI. Input: approval signal. Output: deployed change with CI pipeline running. 8. [TOOL: Hermes Cron] Verification: after deployment, Hermes runs a post-deploy health check cron job that hits the /health endpoint, checks error rates, and sends a deployment summary to the Telegram group. Input: deployment event. Output: health check report with deploy outcome and any anomalies detected.

TOOL INTEGRATION

Hermes Agent v2.0+: Install on a VPS (DigitalOcean $12/month droplet, Ubuntu 22.04) with Docker. Configure the gateway for Telegram using hermes gateway setup. The agent runs as a systemd service with auto-restart. Gotcha: the Hermes MCP server binds to localhost:8642 by default. If Claude Code runs on a different machine, configure hermes mcp set host 0.0.0.0 and add TLS via a reverse proxy (Caddy) to secure the connection. The MCP transport is unencrypted by default. Claude Code Opus 4.8: Install on the same VPS or a separate dev machine. Configure .mcp.json to point at the Hermes MCP server endpoint: {mcpServers: {hermes: {command: npx, args: [@anthropic/mcp-stdio, --url, http://localhost:8642]}}}. Gotcha: Claude Code's MCP client expects a stdio-based transport by default. Use the @anthropic/mcp-stdio bridge package to forward stdio to the Hermes HTTP endpoint. Hermes MCP Server: Built into Hermes Agent v0.4.0+. Enable with hermes mcp enable. The server exposes tools: terminal (execute shell commands), read_file, write_file, web_search (via Tavily or Nous Portal), memory_read/write, skills_list/run, and run_agent (spawn a Hermes subagent). Gotcha: the run_agent tool spawns a Hermes subagent that has full tool access including terminal. Scope the tools available to Claude Code via the MCP server's allowed_tools configuration to prevent Claude Code from spawning recursive agent chains. Telegram Gateway: Hermes uses a polling-based Telegram bot. Configure with BOT_TOKEN from @BotFather. The gateway supports markdown-formatted messages up to 4096 characters. Gotcha: if Claude Code returns a large diff exceeding 4096 characters, configure the gateway message formatter to split output into code blocks of 4000 characters each and send as a message thread. Claude Code Subagents: For complex coding tasks, configure Claude Code to use subagents running in worktrees with --subagent-max-count 3. Each subagent gets its own context window for isolated file editing. Gotcha: subagents in worktrees share the Hermes MCP server connection. If two subagents call run_agent simultaneously, Hermes queues the second call. Set HERMES_MCP_MAX_CONCURRENT=2 in the Hermes environment to prevent resource exhaustion.

ROI METRICS

1. Ops tasks handled without opening a terminal: Before 0% (all ops requires SSH or dashboard login) → After 85% handled through Telegram via Hermes, measured as tasks completed without any SSH session. 2. Code change turnaround time (from request to deployed fix): Before 45 minutes for a simple hotfix (context switch, clone, fix, test, deploy) → After 8-12 minutes (request via Telegram, Claude implements, human reviews, auto-deploys). 3. Monitoring coverage: Before 4 manual health checks per day → After continuous monitoring with 15-minute cron intervals, 3 alert channels (Telegram, email, SMS via gateway). 4. Team capacity multiplier: Before 3 people handling both ops and dev → After effectively 5-person team output with agents handling monitoring, triage, code review, and deployment automation. 5. Incident MTTR (mean time to resolve): Before 2.5 hours for production incidents (detect, page, SSH, investigate, fix, deploy) → After 35 minutes for common failure patterns with automated detection via Hermes monitoring and auto-remediation via Claude Code fixes.

CAVEATS

1. MCP bridge latency: Each cross-agent call through the Hermes MCP server adds 500ms-2s of transport overhead. For time-sensitive incident response, configure Claude Code with a direct SSH tool access as a fallback for the first response while the bridge processes the structured delegation. 2. Gateway message format limitations: Telegram group chats with multiple agents posting can become noisy. Configure separate Telegram topics or channels: one for ops alerts, one for code review summaries, one for deployment notifications. Use the send_message tool's chat_id parameter to route messages per channel. 3. Task routing misclassification: Hermes may route a coding task to itself and attempt file edits through the MCP bridge back to Claude Code, creating a slow round-trip. Train the classifier with 20+ examples per category in the rubric stored in ~/.hermes/profiles/main/rubric.yaml. 4. State divergence: If Claude Code modifies files on disk without updating Hermes's memory, Hermes's project context becomes stale. Configure a post-edit hook in Claude Code that sends a memory_write call to the Hermes MCP server after significant file changes.