Folders and files
| Name | Name | Last commit date | ||
|---|---|---|---|---|
Local-first agent dispatch for the Hermes AI harness. Route and expand prompts across 28 specialized agents, reachable from your phone over Tailscale.
There's no framework here and no hand-rolled agent loop. Hermes is the loop. Each agent is the harness constrained by a few per-agent flags plus one context-minimal project config, wrapped as a pure stdin → stdout shell script. The model you route to is the quality lever: cheap models for triage and extraction, one big model reserved for the work that's actually hard.
- Runs 28 narrow agents. Each is a system prompt plus a model alias. They cover writing, code review, SEO, GTM, sales, legal, finance, knowledge work, and automation.
- Gives you one chat box on your phone. A fast model picks the right agent and rewrites your terse message into a tuned prompt, then the chosen agent runs and streams back.
- Stays portable. Every agent names an alias (max, code, structured), never a model. Point those 7 aliases at MLX, CUDA, cloud, or a single model. Same agents, any backend.
- Hermes Agent v0.15+ on your PATH.
- A model backend: Ollama for local models, LiteLLM for the proxy and cloud. The path is hermes → LiteLLM (:4000) → backend.
- jq for pipelines, Python 3.9+ for the dispatch server (stdlib only).
- Tailscale if you want phone access. Optional.
| Path | For | Start here |
|---|---|---|
| DIY | "I want full control, I'll wire it myself." | docs/diy.md |
| Do it With Me | "Walk me through it." | ./setup.sh, then docs/diwm.md |
| Do it For Me | "Just run it in containers." | docs/difm.md (docker/) |
git clone <this-repo> hermes-dispatch && cd hermes-dispatch
cp config.example.yaml config.yaml # then edit, or:
./setup.sh # interactive wizard (DIWM)Every agent maps to one of 7 aliases: three capability tiers (fast, balanced, max) and four task roles (structured, code, writing, reasoning) that default to a tier. What backs each is your choice. See docs/model-configuration.md.
| Agent | Purpose | Alias |
|---|---|---|
| blog-drafter | First-draft long-form blog prose from an outline | writing |
| social-media-marketer | Topic or draft to platform-tailored LinkedIn and X posts | fast |
| status-update-writer | done/blocked/next bullets to a client status update | writing |
| devrel-sample | API plus capability description to a developer code sample | code |
| pr-reviewer | Unified diff to a five-axis code review | max |
| seo-reviewer | Page or draft to a qualitative on-page SEO and AEO audit | max |
| seo-tester | Page to machine-readable pass/fail SEO checks (JSON) | structured |
| gtm-planner | Product brief to a decision-ready go-to-market plan | max |
| gtm-executor | GTM plan to paste-ready launch assets | writing |
| lead-designer | Brief to design direction or critique (text specs) | max |
| prospect-researcher | Company or contact to a sales research brief | max |
| discovery-prep | Prospect plus goal to a pre-call discovery plan | max |
| proposal-writer | Scope brief to a structured client proposal | writing |
| followup-drafter | Last interaction to a follow-up email | writing |
| scope-creep-detector | Agreed scope plus new request to an in/out call | reasoning |
| contract-reviewer | Contract or vendor agreement to a severity-ranked risk pass | max |
| tos-reviewer | Terms of service to a risk pass before you agree | max |
| privacy-checker | Privacy policy or DPA to a data map plus gaps | max |
| cashflow-summarizer | Transactions to a weekly cashflow summary | reasoning |
| invoice-tracker | Invoice email to vendor, amount, due date, and status | structured |
| expense-classifier | Transaction descriptions to tax-relevant categories | structured |
| triage-router | One item to a routing decision (minified JSON) | structured |
| inbox-triage | Batch of emails to a classified enum plus urgency | structured |
| vault-distiller | Long note to atomic concepts, [[wikilinks]], and tags | max |
| decision-journal | A decision to a durable note written for future-you | reasoning |
| meeting-to-actions | Notes or transcript to decisions plus action items | reasoning |
| retro-generator | Project notes to a structured retrospective | reasoning |
| weekly-review-synthesis | A week of notes to a candid weekly review | max |
Run one directly:
git diff main... | ./agents/pr-reviewer/run.sh
./agents/triage-router/run.sh "We need to launch the toolkit next week"
HERMES_DRY_RUN=1 ./agents/blog-drafter/run.sh "test" # inspect the composed command- Chain agents into pipelines over a shared file-based state store: bin/run-pipeline.sh gtm-pipeline "Launch brief: ...". See pipelines/.
- Expose every agent as a desktop chat persona: bin/gen-profiles.sh, then hermes desktop. See docs/hermes-desktop.md.
- Reach the roster from your phone: python3 dispatch/server.py plus Tailscale. See docs/tailscale-mobile.md.
| Doc | What it covers |
|---|---|
| docs/quickstart.md | Fastest path from clone to first agent reply |
| docs/model-configuration.md | How aliases work, what backs them, hardware tradeoffs |
| docs/agents.md | Full agent roster with descriptions and recommended aliases |
| docs/hermes-desktop.md | Profile registration, desktop and dashboard visibility |
| docs/tailscale-mobile.md | Tailscale Serve setup for phone access |
| docs/obsidian.md | Optional Obsidian save-to-vault integration |
| docs/diy.md, docs/diwm.md, docs/difm.md | The three setup tiers |
MIT.