First agency
Tools, spend control, and a multi-step proof on Desktop after the gateway works.
Agent with tools
Telegram already reaches the agent. Now enable hands: terminal, browser, filesystem, search when configured. The model requests; Hermes executes.
In Desktop: open Tools / Dashboard, confirm toolsets for this profile, run the proof from Desktop chat. Telegram remains a second surface.
- ·Prefer structured tools over long shell improvisation when both exist.
- ·A failing tool should return a clear error so the model can recover.
- ·Scope tools to the job — research often needs read and web, not full shell write.
- ·If Portal enabled Tool Gateway tools, verify them in the Tools UI.
- ·Resume sessions for long jobs when the UI offers it.
Terminal backend (Advanced — when tools hit the shell)
Shell and file tools use a terminal backend. Day one stays local (host process). Settings → Advanced may offer Docker and other backends so commands run with isolation from the host.
- ·Local (course default) — simplest; approvals still apply
- ·Docker — isolation; enable only after chat and tools work
- ·SSH / cloud sandboxes — remote workers for later
MCP — connect services carefully
MCP connects external services (docs, drives, repos, APIs) through a standard interface. Treat each server as new code on the trust surface.
- ·Add one server for a concrete need.
- ·Filter tools to the smallest useful surface.
- ·Prefer OAuth over long-lived tokens when available.
- ·Keep secrets out of chat; store them in .env / setup.
Spend control
Provider and model are separate decisions. Provider is auth and billing; model is capability and price. Use a strong model where mistakes are expensive; cheaper models for auxiliary jobs when supported.
- ·Stabilize one primary model before fallbacks.
- ·Credential pools help with rate limits, not only multi-billing.
- ·Local models: set sufficient context; broken tool loops are often truncated context.
- ·Providers is also where remote or local OpenAI-compatible endpoints appear.
Fixed cost of an empty session
Before the first user message, Hermes already loads system prompt, skills index, MEMORY/USER snapshots, tool schemas, and project context (for example AGENTS.md). That fixed budget is often why an agent feels slow or expensive — not a “weak model.”
Session commands (map)
Long sessions fill the context window. Official and community practice uses in-session commands rather than starting over every time.
- ·/context — what consumes the window and how much room remains
- ·/compress — summarize older turns; keep recent exchanges or focus a topic (session file stays)
- ·/undo — rewind last user turn(s) into the composer; does not restore files already changed
- ·/retry · /branch — same prompt again, or explore another path without losing the first
- ·/rollback + checkpoints — restore files when the agent already wrote to disk
Proof task
Desktop cockpit
If you followed module 03, Hermes Desktop is the everyday surface for tools, approvals, and sessions for the rest of the course.
Optional stretch: request a small custom Desktop widget if your build supports it (for example a memory monitor). Agency on phone and machine is the goal.
Human in the loop
Dangerous actions require approval. Smart mode may pass low-risk actions and escalate uncertainty. Keep approvals on for this course. YOLO-style modes are expert-only and still cannot bypass hardline host protections — do not treat them as safe defaults.
After this proof
Tools and spend are introduced here. To measure one job and reduce a toolset, use the spend drill lab — not a second lecture.
A multi-step tool task left a file or on-screen result you can open. You know where model cost is decided (provider vs model).
Stored on this device only — not sent anywhere.
End of Part I
Part II starts next: memory, vault, skills, security, and cron — the compounding harness.