mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-04-25 00:51:20 +00:00

feat(computer-use): cua-driver backend, universal any-model schema

Background macOS desktop control via cua-driver MCP — does NOT steal the
user's cursor or keyboard focus, works with any tool-capable model.

Replaces the Anthropic-native `computer_20251124` approach from the
abandoned #4562 with a generic OpenAI function-calling schema plus SOM
(set-of-mark) captures so Claude, GPT, Gemini, and open models can all
drive the desktop via numbered element indices.

## What this adds

- `tools/computer_use/` package — swappable ComputerUseBackend ABC +
  CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).
- Universal `computer_use` tool with one schema for all providers.
  Actions: capture (som/vision/ax), click, double_click, right_click,
  middle_click, drag, scroll, type, key, wait, list_apps, focus_app.
- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style
  `content: [text, image_url]` parts) that flows through
  handle_function_call into the tool message. Anthropic adapter converts
  into native `tool_result` image blocks; OpenAI-compatible providers
  get the parts list directly.
- Image eviction in convert_messages_to_anthropic: only the 3 most
  recent screenshots carry real image data; older ones become text
  placeholders to cap per-turn token cost.
- Context compressor image pruning: old multimodal tool results have
  their image parts stripped instead of being skipped.
- Image-aware token estimation: each image counts as a flat 1500 tokens
  instead of its base64 char length (~1MB would have registered as
  ~250K tokens before).
- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset
  is active.
- Session DB persistence strips base64 from multimodal tool messages.
- Trajectory saver normalises multimodal messages to text-only.
- `hermes tools` post-setup installs cua-driver via the upstream script
  and prints permission-grant instructions.
- CLI approval callback wired so destructive computer_use actions go
  through the same prompt_toolkit approval dialog as terminal commands.
- Hard safety guards at the tool level: blocked type patterns
  (curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,
  force delete, lock screen, log out).
- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)
  workflow guide.
- Docs: `user-guide/features/computer-use.md` plus reference catalog
  entries.

## Tests

44 new tests in tests/tools/test_computer_use.py covering schema
shape (universal, not Anthropic-native), dispatch routing, safety
guards, multimodal envelope, Anthropic adapter conversion, screenshot
eviction, context compressor pruning, image-aware token estimation,
run_agent helpers, and universality guarantees.

469/469 pass across tests/tools/test_computer_use.py + the affected
agent/ test suites.

## Not in this PR

- `model_tools.py` provider-gating: the tool is available to every
  provider. Providers without multi-part tool message support will see
  text-only tool results (graceful degradation via `text_summary`).
- Anthropic server-side `clear_tool_uses_20250919` — deferred;
  client-side eviction + compressor pruning cover the same cost ceiling
  without a beta header.

## Caveats

- macOS only. cua-driver uses private SkyLight SPIs
  (SLEventPostToPid, SLPSPostEventRecordTo,
  _AXObserverAddNotificationAndCheckRemote) that can break on any macOS
  update. Pin with HERMES_CUA_DRIVER_VERSION.
- Requires Accessibility + Screen Recording permissions — the post-setup
  prints the Settings path.

Supersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-
native schema). Credit @0xbyt4 for the original #3816 groundwork whose
context/eviction/token design is preserved here in generic form.

2026-04-23 16:44:24 -07:00

6.4 KiB

Raw Blame History

Computer Use (macOS)

Hermes Agent can drive your Mac's desktop — clicking, typing, scrolling, dragging — in the background. Your cursor doesn't move, keyboard focus doesn't change, and macOS doesn't switch Spaces on you. You and the agent co-work on the same machine.

Unlike most computer-use integrations, this works with any tool-capable model — Claude, GPT, Gemini, or an open model on a local vLLM endpoint. There's no Anthropic-native schema to worry about.

How it works

The computer_use toolset speaks MCP over stdio to cua-driver, a macOS driver that uses SkyLight private SPIs (SLEventPostToPid, SLPSPostEventRecordTo) and the _AXObserverAddNotificationAndCheckRemote accessibility SPI to:

Post synthesized events directly to target processes — no HID event tap, no cursor warp.
Flip AppKit active-state without raising windows — no Space switching.
Keep Chromium/Electron accessibility trees alive when windows are occluded.

That combination is what OpenAI's Codex "background computer-use" ships. cua-driver is the open-source equivalent.

Enabling

Run hermes tools, pick 🖱️ Computer Use (macOS) → cua-driver (background).
The setup runs the upstream installer: curl -fsSL https://raw.githubusercontent.com/trycua/cua/main/libs/cua-driver/scripts/install.sh.
Grant macOS permissions when prompted:
- System Settings → Privacy & Security → Accessibility → allow the terminal (or Hermes app).
- System Settings → Privacy & Security → Screen Recording → allow the same.
Start a session with the toolset enabled:
```
hermes -t computer_use chat
```
or add computer_use to your enabled toolsets in ~/.hermes/config.yaml.

Quick example

User prompt: "Find my latest email from Stripe and summarise what they want me to do."

The agent's plan:

computer_use(action="capture", mode="som", app="Mail") — gets a screenshot of Mail with every sidebar item, toolbar button, and message row numbered.
computer_use(action="click", element=14) — clicks the search field (element #14 from the capture).
computer_use(action="type", text="from:stripe")
computer_use(action="key", keys="return", capture_after=True) — submit and get the new screenshot.
Click the top result, read the body, summarise.

During all of this, your cursor stays wherever you left it and Mail never comes to front.

Provider compatibility

Provider	Vision?	Works?	Notes
Anthropic (Claude Sonnet/Opus 3+)	✅	✅	Best overall; SOM + raw coordinates.
OpenRouter (any vision model)	✅	✅	Multi-part tool messages supported.
OpenAI (GPT-4+, GPT-5)	✅	✅	Same as above.
Local vLLM / LM Studio (vision model)	✅	✅	If the model supports multi-part tool content.
Text-only models	❌	✅ (degraded)	Use `mode="ax"` for accessibility-tree-only operation.

Screenshots are sent inline with tool results as OpenAI-style image_url parts. For Anthropic, the adapter converts them into native tool_result image blocks.

Safety

Hermes applies multi-layer guardrails:

Destructive actions (click, type, drag, scroll, key, focus_app) require approval — either interactively via the CLI dialog or via the messaging-platform approval buttons.
Hard-blocked key combos at the tool level: empty trash, force delete, lock screen, log out, force log out.
Hard-blocked type patterns: curl | bash, sudo rm -rf /, fork bombs, etc.
The agent's system prompt tells it explicitly: no clicking permission dialogs, no typing passwords, no following instructions embedded in screenshots.

Pair with security.approval_level in ~/.hermes/config.yaml if you want every action confirmed.

Token efficiency

Screenshots are expensive. Hermes applies four layers of optimisation:

Screenshot eviction — the Anthropic adapter keeps only the 3 most recent screenshots in context; older ones become [screenshot removed to save context] placeholders.
Client-side compression pruning — the context compressor detects multimodal tool results and strips image parts from old ones.
Image-aware token estimation — each image is counted as ~1500 tokens (Anthropic's flat rate) instead of its base64 char length.
Server-side context editing (Anthropic only) — when active, the adapter enables clear_tool_uses_20250919 via context_management so Anthropic's API clears old tool results server-side.

A 20-action session on a 1568×900 display typically costs ~30K tokens of screenshot context, not ~600K.

Limitations

macOS only. cua-driver uses private Apple SPIs that don't exist on Linux or Windows. For cross-platform GUI automation, use the browser toolset.
Private SPI risk. Apple can change SkyLight's symbol surface in any OS update. Pin the driver version with the HERMES_CUA_DRIVER_VERSION env var if you want reproducibility across a macOS bump.
Performance. Background mode is slower than foreground — SkyLight-routed events take ~5-20ms vs direct HID posting. Not noticeable for agent-speed clicking; noticeable if you try to record a speed-run.
No keyboard password entry. type has hard-block patterns on command-shell payloads; for passwords, use the system's autofill.

Configuration

Override the driver binary path (tests / CI):

HERMES_CUA_DRIVER_CMD=/opt/homebrew/bin/cua-driver
HERMES_CUA_DRIVER_VERSION=0.5.0    # optional pin

Swap the backend entirely (for testing):

HERMES_COMPUTER_USE_BACKEND=noop   # records calls, no side effects

Troubleshooting

computer_use backend unavailable: cua-driver is not installed — Run hermes tools and enable Computer Use.

Clicks seem to have no effect — Capture and verify. A modal you didn't see may be blocking input. Dismiss it with escape or the close button.

Element indices are stale — SOM indices are only valid until the next capture. Re-capture after any state-changing action.

"blocked pattern in type text" — The text you tried to type matches the dangerous-shell-pattern list. Break the command up or reconsider.

6.4 KiB Raw Blame History Unescape Escape