A hand-written Linux daemon unit (systemd user unit / XDG autostart entry
running `cua-driver serve`) can be dead for days — crash loop, stopped, never
started — while `hermes computer-use doctor` and `status` report a healthy
binary: the runtime contract only checks the binary (`manifest`), and nothing
ever connected to the daemon socket (#114748).
- tools/computer_use/cua_backend.py::cua_daemon_listening — socket-level
liveness via `cua-driver status [--socket PATH]` (rc 0 = a daemon answered,
"not running" = dead, anything else = unknown). Never raises.
- tools/computer_use/doctor.py::cua_daemon_units — one scan of the units that
run cua-driver (kind, unit, exec target, runs `serve`, `--socket` path with
`%h` expanded); the pruned-Exec guard now filters that list instead of
re-scanning.
- doctor.py::_apply_daemon_liveness_guard — per `serve` unit: `pass` when its
socket answers, `fail` (degrading `ok`) when not, with the hint that a driver
reinstall does not start the daemon. Unconfigured daemons are never probed:
on Linux the MCP runtime needs none, so a silent default socket is normal.
- `hermes computer-use status` prints the dead-daemon line and exits 1.
- docs: the Linux daemon-unit paragraph under the doctor section.
Not changed: _maybe_repair_runtime_contract. The repair is gated on the
binary-level contract only; a dead daemon never enters it, so a reinstall was
never triggered by the daemon (the `.release_installed/<version>` marker the
reporter saw is written by cua-driver itself on any first run of the binary —
observed via strace of `cua-driver status` under a fresh HOME).
Supersedes #114928 (@Finn763): same finding, ~600-line implementation with a
new module and a repair-gate rewrite; this is the ~90-line version on the
existing doctor seams.
The skill described a SOM overlay burned into the screenshot, a driver-side
`capture` tool, and manual symlinking of the cua-driver skill pack; users
who read the driver's own docs then called raw MCP tools (`capture`,
bare `element_index`) and hit "no reviewed risk classification" and
`snapshot_id_required`. State plainly that `computer_use(action=...)` is a
wrapper vocabulary the driver never sees, that `element=N` is translated to
the snapshot token, what a `stale` refusal means, how text-only models get
vision (auxiliary.vision routing / mode=ax), and the Windows WindowsApps
doctor failure. `cua-driver skills install` links into ~/.hermes/skills now.
computer_use kept its own approval decision: two module dicts
(_session_auto_approve / _always_allow) mirroring tools.approval's
session store and _persist_choice, a private verdict vocabulary
(approve_once/approve_session/always_approve) that hermes_cli mapped
back to once/session/always, and — the real problem — `if
_approval_callback is None: return None`. Only the interactive CLI ever
installed that callback, so every other host (gateway turns, cron,
api_server, tui_gateway, ACP) ran destructive desktop input with no
approval at all, ignoring cron_mode / unattended_mode / the permanent
allowlist, and "always" grants were invisible to `is_approved`,
`clear_session` and the messaging-platform approval buttons.
_request_approval now calls tools.approval._run_approval_gate with
pattern_key `cua:<action>:<background|foreground>` (the old scope shape,
so a background grant still never covers the visible foreground variant)
and fail_closed_when_no_human=True, the same posture as
request_tool_approval / the SSH-config write gate. The private dicts,
their release/atexit clearing, the verdict mapping in
hermes_cli/cli_modal_mixin.py and the extra callback install in cli.py
are deleted: the CLI's terminal_tool callback answers computer_use
prompts like any other tool. set_approval_callback stays as an optional
explicit-callback hook with the shared callback contract
(cb(command, description, **kw) -> once|session|always|deny|timeout);
no in-tree host uses it.
Behavior change:
- No approval callback and no gateway (cron, api_server/webhook,
headless -q, plain library use): destructive actions are now REFUSED
with a BLOCKED error and never reach the backend. Previously they
silently ran. cron honors approvals.cron_mode, unattended platforms
approvals.unattended_mode, -q approvals.single_query_mode.
- --yolo / gateway /yolo / approvals.mode: off still allow (unchanged).
- Gateway sessions (Telegram/Discord/Slack/...) now get a real pending
approval with once/session/always buttons instead of default-allow.
- session/always grants live in tools.approval's store; "always" is one
command_allowlist entry (`cua:click:background`) and is scoped to that
action+mode — the old blanket "always_approve unlocks everything for
the session" no longer exists.
- Denial wording is the shared gate's ("BLOCKED: User denied ...",
"BLOCKED: Action timed out ..."); the error JSON keeps `action`.
Tests: tests/tools/test_computer_use_approval_isolation.py
::test_no_callback_refuses_unless_yolo (blocked + no backend call, then
yolo executes) and ::test_always_grant_lands_in_the_shared_store
(is_approved sees the cua:<action>:<mode> key; second call served from
the store). Sabotage: restoring the `callback is None -> allow`
short-circuit fails the first; swapping the shared gate for a private
grant set fails the second plus the three delivery-ladder scope tests.
tests/tools/conftest.py gains `grant_computer_use_approvals` for
dispatch tests that only care about routing.
The grant_existing_profile key was removed in PR #98057; sweep the mode
table, opt-in section, runtime-lifecycle notes, config example, and CLI
reference that still documented it.
#95620 removed computer_use's cua_browser_* actions entirely
(browser_route.py, the action enum, and their dispatch/escalation hint) —
computer_use is desktop-only now, and page content goes through the
separate browser_navigate/browser_click/... toolset (or browser_exec
under the Browser Use CLI backend).
The skill doc never caught up: it still taught the model to call
cua_browser_state/cua_browser_prepare/etc. and to escalate to a "page"
rung that _enrich_escalation can no longer recommend. Following that
guidance fails schema validation on the first call.
- SKILL.md: replace the "Typed browser page rung" section (including the
existing_profile authorization walkthrough, which described a
model-facing action parameter that no longer exists anywhere in
schema.py/tool.py) with a short pointer to the current browser toolset;
drop "page" from the escalation.recommended union and the ladder step
that referenced it.
- website/docs/user-guide/features/computer-use.md: grant_existing_profile
and the rest of the permission-mode config are still real and current
(cua_backend.py still reads them for the runtime launch grant) — only
reworded the one sentence naming cua_browser_prepare as the mechanism
that consumes the grant, since driving a signed-in browser window now
goes through the standard capture/click/type actions instead.
- Regenerated the mirrored skill doc via
website/scripts/generate-skill-docs.py rather than hand-editing it, per
its own header. Kept the diff scoped to computer-use only.
Full-screen captures carry no element tree, so the CaptureResult now has a
'note' field surfaced in the tool summary telling the model to call
capture(app='<AppName>') or capture(app='desktop') when it needs to act on
what it sees. Schema description updated to distinguish app='screen'
(composited full-screen image) from app='desktop' (shell surface with
clickable elements); docs + regression tests (14, sabotage-verified) added.
Follow-up to PR #90183 — computer_use now saves a bounded shareable copy of
image captures, so attachment-capable surfaces (Telegram, Discord, Desktop)
can deliver the real screenshot when the user asks. Documents the behavior,
the 20-file cache bound, and the no-automatic-send rule.
The runtime-contract repair now also runs during hermes update and once
per session at the first computer_use call (PR #87923); the docs only
mentioned setup and toolset enablement.
Live-tested against the real cua-driver 0.19.3 binary (Linux x86_64):
- bounded serve flags corrected: the daemon accepts
--session-policy/--approve-session-policy, not the docs'
--capability-manifest names (which it rejects). Verified end-to-end:
a bounded daemon with a real policy file starts and reports running.
- browser-approve verified real but interactive-only (refuses without a
TTY) and its token is a legacy compatibility path disabled by default
on current drivers (per the live browser_prepare schema). Kept as a
passthrough; no longer presented as the primary route.
- NEW primary standard-mode route, verified live: launch the runtime
with cua-driver's trusted-launcher grant. config opt-in
computer_use.grant_existing_profile: true appends
--grant existing-profile to the standard-mode MCP spawn (MCP
initialize verified accepting the flag). Default false = attachment
keeps failing closed. Never applied to bounded/unrestricted daemons.
- Skill, system prompt, tool schema, and docs updated to the verified
ladder: config grant > bounded manifest > YOLO; token = legacy.
Completes the typed cua_browser_* route (PR #74166 lineage) with the
authorization surface that makes existing-profile attachment and
repeatable bounded automation reachable by real users:
- hermes computer-use browser-approve: CLI passthrough that mints
cua-driver's five-minute single-use attachment token for one exact
(pid, window_id). The user, never the model, is the token source.
- approval_token passthrough on cua_browser_prepare (schema + dispatch +
browser_route), forwarded only for existing_profile and only as a
non-empty string.
- computer_use.permission_mode: bounded + capability_manifest config:
private per-session embedded daemon launched with
--capability-manifest/--approve-capability-manifest; missing manifest
fails loudly. 'unrestricted' is deliberately NOT a config value —
it stays bound to the explicit per-session YOLO toggle.
- Skill + system-prompt + docs guidance for the three authorization
rungs and the isolated-profile-first default.
E2E-verified against a temp HERMES_HOME: real config resolution to
bounded, loud failure without a manifest, real argparse path driving a
fake cua-driver binary, standard default preserved.
Choosing Computer Use should be a config flip, not a hunt for
'hermes computer-use install'. Three provisioning rungs:
- install.sh / install.ps1 pre-install cua-driver (best-effort,
non-fatal, time-boxed at 660s above the upstream installer's 600s
lock window; --skip-computer-use / -SkipComputerUse to opt out;
Termux and unwritable-/Applications skipped cleanly)
- PUT /api/tools/toolsets/{name} (dashboard + desktop toggle) spawns
the background 'hermes tools post-setup cua_driver' action when the
toolset is enabled while the binary is missing — previously the
toggle 'saved' but the tool never appeared in the schema because
check_computer_use_requirements() couldn't find the binary
- hermes tools interactive flow already installed via
_toolset_needs_configuration_prompt/_POST_SETUP_INSTALLED (unchanged)
Docs: computer-use.md enabling section rewritten around the new flow;
installation.md documents --skip-computer-use.
skills/computer-use/SKILL.md sat at the root of the bundled skills
tree as an uncategorized single-skill directory. Move it to
skills/autonomous-ai-agents/computer-use/ alongside claude-code,
codex, opencode, and hermes-agent.
- git mv skills/computer-use -> skills/autonomous-ai-agents/computer-use
(history preserved)
- root-level-skill docstring examples (commands.py,
generate-skill-docs.py) now use a generic placeholder instead of
naming a specific skill, since categorizing root-level skills is
ongoing
- docs: move the bundled page to autonomous-ai-agents-computer-use,
update sidebars.ts, skills-catalog.md, and the two skill-path
references in features/computer-use.md
E2E validated: fresh sync_skills() copies to the new nested path;
existing installs with the old flat copy keep it untouched (manifest
name-keyed, hash unchanged -> skipped, no duplicate);
_get_category_from_path resolves 'autonomous-ai-agents'; docusaurus
build green; 396 targeted tests pass.
A Medium-integrity Hermes agent cannot drive High-integrity (admin)
windows on Windows — UIPI blocks UIA enumeration and mouse injection
(SOM returns 0 elements, clicks silently no-op, screenshots still work,
keyboard partially bypasses). OS constraint affecting every Windows
automation stack, not a cua-driver bug. Document the symptom + the
run-elevated workaround. Closes#49067.
* feat(computer_use): disable cua-driver telemetry by default, add opt-in
cua-driver ships anonymous PostHog usage telemetry ENABLED by default
upstream (fires cua_driver_install / cua_driver_doctor events to
eu.i.posthog.com). Hermes now disables it for our users unless they
explicitly opt in.
- New config key `computer_use.cua_telemetry` (default false) in
DEFAULT_CONFIG.
- `cua_backend.cua_driver_child_env()` injects
`CUA_DRIVER_RS_TELEMETRY_ENABLED=0` into the child env when telemetry is
disabled (the default); leaves the var untouched on opt-in so the driver
uses its own default. Reads config fail-safe — any error defaults to
telemetry off.
- Routed every cua-driver spawn site through the policy: MCP backend
(StdioServerParameters env), `cua_driver_update_check`, doctor's
health_report Popen, the install.sh/install.ps1 runner, and the
`--version` / status probes.
- Docs: new Telemetry subsection in computer-use.md (EN).
- Tests: tests/computer_use/test_cua_telemetry.py — default disables,
explicit-false disables, opt-in leaves var untouched, config-failure
fails safe, inherited-enabled is overridden off.
Verified live on Linux against the real cua-driver-rs 0.6.0 binary: with
the var=0 the driver reports "telemetry: disabled via
CUA_DRIVER_RS_TELEMETRY_ENABLED" and sends no event; with it unset it logs
"sending event: cua_driver_doctor". 213 computer_use + install tests green.
* fix(dashboard): fold computer_use config category into agent tab
The new computer_use.cua_telemetry key created a single-field dashboard
config category, tripping test_no_single_field_categories (web_server's
invariant that categories with <2 fields must be merged to avoid tab
sprawl). Add computer_use -> agent to _CATEGORY_MERGE, matching the
existing onboarding/telegram single-field folds.
Make the computer_use toolset platform-agnostic by driving cua-driver on
macOS, Windows, and Linux. Consumes the 8 cua-driver decoupling surfaces
(capability discovery, structuredContent AX tree, opaque element_token,
click button enum, explicit mimeType, machine-readable manifest,
structured list_windows, structured health_report), each degrading
gracefully on older drivers.
Adds `hermes computer-use doctor` (drives cua-driver health_report with a
per-OS check matrix and an exit 0/1/2 ok/degraded/blocked contract), full
typed wrappers for the previously-uncovered cua-driver tools plus a generic
call_tool escape hatch, per-session agent-cursor lifecycle, platform-aware
system-prompt guidance (host-deterministic, cache-safe), and honors
HERMES_CUA_DRIVER_CMD end-to-end.
Replaces the macOS-only skills/apple/macos-computer-use skill with a
cross-platform skills/computer-use skill, and refreshes the EN + zh-Hans
docs.
Supersedes #44221 (Windows-enablement salvage of #30660).
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
cua-driver was only installed once on toolset enable: `_run_post_setup` early-returns when the binary is already on PATH, so upstream fixes (e.g. v0.1.6 Safari window-focus fix) never reached existing users without manual reinstall.
Two refresh points now:
- `hermes update` re-runs the upstream installer at the end of the update if cua-driver is on PATH (macOS-only, no-op otherwise). Ties driver freshness to the user-controlled update cadence — no startup latency, no per-launch GitHub API call.
- `hermes computer-use install --upgrade` for manual force-refresh.
The upstream `install.sh` always pulls the latest release, so re-running is the canonical upgrade path. No version-comparison logic needed.
`hermes computer-use status` now shows the installed version, and points at `--upgrade` for refreshing.
* docs: deep audit — fix stale config keys, missing commands, and registry drift
Cross-checked ~80 high-impact docs pages (getting-started, reference, top-level
user-guide, user-guide/features) against the live registries:
hermes_cli/commands.py COMMAND_REGISTRY (slash commands)
hermes_cli/auth.py PROVIDER_REGISTRY (providers)
hermes_cli/config.py DEFAULT_CONFIG (config keys)
toolsets.py TOOLSETS (toolsets)
tools/registry.py get_all_tool_names() (tools)
python -m hermes_cli.main <subcmd> --help (CLI args)
reference/
- cli-commands.md: drop duplicate hermes fallback row + duplicate section,
add stepfun/lmstudio to --provider enum, expand auth/mcp/curator subcommand
lists to match --help output (status/logout/spotify, login, archive/prune/
list-archived).
- slash-commands.md: add missing /sessions and /reload-skills entries +
correct the cross-platform Notes line.
- tools-reference.md: drop bogus '68 tools' headline, drop fictional
'browser-cdp toolset' (these tools live in 'browser' and are runtime-gated),
add missing 'kanban' and 'video' toolset sections, fix MCP example to use
the real mcp_<server>_<tool> prefix.
- toolsets-reference.md: list browser_cdp/browser_dialog inside the 'browser'
row, add missing 'kanban' and 'video' toolset rows, drop the stale
'38 tools' count for hermes-cli.
- profile-commands.md: add missing install/update/info subcommands, document
fish completion.
- environment-variables.md: dedupe GMI_API_KEY/GMI_BASE_URL rows (kept the
one with the correct gmi-serving.com default).
- faq.md: Anthropic/Google/OpenAI examples — direct providers exist (not just
via OpenRouter), refresh the OpenAI model list.
getting-started/
- installation.md: PortableGit (not MinGit) is what the Windows installer
fetches; document the 32-bit MinGit fallback.
- installation.md / termux.md: installer prefers .[termux-all] then falls
back to .[termux].
- nix-setup.md: Python 3.12 (not 3.11), Node.js 22 (not 20); fix invalid
'nix flake update --flake' invocation.
- updating.md: 'hermes backup restore --state pre-update' doesn't exist —
point at the snapshot/quick-snapshot flow; correct config key
'updates.pre_update_backup' (was 'update.backup').
user-guide/
- configuration.md: api_max_retries default 3 (not 2); display.runtime_footer
is the real key (not display.runtime_metadata_footer); checkpoints defaults
enabled=false / max_snapshots=20 (not true / 50).
- configuring-models.md: 'hermes model list' / 'hermes model set ...' don't
exist — hermes model is interactive only.
- tui.md: busy_indicator -> tui_status_indicator with values
kaomoji|emoji|unicode|ascii (not kawaii|minimal|dots|wings|none).
- security.md: SSH backend keys (TERMINAL_SSH_HOST/USER/KEY) live in .env,
not config.yaml.
- windows-wsl-quickstart.md: there is no 'hermes api' subcommand — the
OpenAI-compatible API server runs inside hermes gateway.
user-guide/features/
- computer-use.md: approvals.mode (not security.approval_level); fix broken
./browser-use.md link to ./browser.md.
- fallback-providers.md: top-level fallback_providers (not
model.fallback_providers); the picker is subcommand-based, not modal.
- api-server.md: API_SERVER_* are env vars — write to per-profile .env,
not 'hermes config set' which targets YAML.
- web-search.md: drop web_crawl as a registered tool (it isn't); deep-crawl
modes are exposed through web_extract.
- kanban.md: failure_limit default is 2, not '~5'.
- plugins.md: drop hard-coded '33 providers' count.
- honcho.md: fix unclosed quote in echo HONCHO_API_KEY snippet; document
that 'hermes honcho' subcommand is gated on memory.provider=honcho;
reconcile subcommand list with actual --help output.
- memory-providers.md: legacy 'hermes honcho setup' redirect documented.
Verified via 'npm run build' — site builds cleanly; broken-link count went
from 149 to 146 (no regressions, fixed a few in passing).
* docs: round 2 audit fixes + regenerate skill catalogs
Follow-up to the previous commit on this branch:
Round 2 manual fixes:
- quickstart.md: KIMI_CODING_API_KEY mentioned alongside KIMI_API_KEY;
voice-mode and ACP install commands rewritten — bare 'pip install ...'
doesn't work for curl-installed setups (no pip on PATH, not in repo
dir); replaced with 'cd ~/.hermes/hermes-agent && uv pip install -e
".[voice]"'. ACP already ships in [all] so the curl install includes it.
- cli.md / configuration.md: 'auxiliary.compression.model' shown as
'google/gemini-3-flash-preview' (the doc's own claimed default);
actual default is empty (= use main model). Reworded as 'leave empty
(default) or pin a cheap model'.
- built-in-plugins.md: added the bundled 'kanban/dashboard' plugin row
that was missing from the table.
Regenerated skill catalogs:
- ran website/scripts/generate-skill-docs.py to refresh all 163 per-skill
pages and both reference catalogs (skills-catalog.md,
optional-skills-catalog.md). This adds the entries that were genuinely
missing — productivity/teams-meeting-pipeline (bundled),
optional/finance/* (entire category — 7 skills:
3-statement-model, comps-analysis, dcf-model, excel-author, lbo-model,
merger-model, pptx-author), creative/hyperframes,
creative/kanban-video-orchestrator, devops/watchers,
productivity/shop-app, research/searxng-search,
apple/macos-computer-use — and rewrites every other per-skill page from
the current SKILL.md. Most diffs are tiny (one line of refreshed
metadata).
Validation:
- 'npm run build' succeeded.
- Broken-link count moved 146 -> 155 — the +9 are zh-Hans translation
shells that lag every newly-added skill page (pre-existing pattern).
No regressions on any en/ page.
Returning users who enabled '🖱️ Computer Use (macOS)' via 'hermes tools'
saw '✓ Saved configuration' but no install — cua-driver was never on
PATH and the toolset failed at first use. Two compounding causes:
1. _toolset_needs_configuration_prompt fell through to _toolset_has_keys,
which returned True for any provider with empty env_vars. cua-driver
has no env vars, so the gate skipped _configure_toolset entirely and
_run_post_setup('cua_driver') never ran.
2. No stable CLI entry-point existed for re-running the install when
the picker no-op'd it (e.g. when toggling the toolset off+on inside
one picker session, where 'added' is empty).
Changes:
- hermes_cli/tools_config.py: add _POST_SETUP_INSTALLED registry
mapping post_setup keys to installed-state predicates. The gate
now returns True when any visible provider has a registered
post_setup whose predicate fails. cua_driver is the only opt-in
for now; other post_setup hooks keep their existing behaviour.
- hermes_cli/main.py: add 'hermes computer-use install' and
'hermes computer-use status' as a stable docs target. install
reuses the same _run_post_setup('cua_driver') path that the
picker invokes; status reports whether cua-driver is on PATH.
- tools/computer_use/cua_backend.py: install hint now points users
at 'hermes computer-use install' first.
- website/docs/user-guide/features/computer-use.md: document the
new command as the primary install path.
- website/docs/reference/cli-commands.md: catalog 'hermes
computer-use' alongside 'hermes tools'.
- tests/hermes_cli/test_post_setup_gating.py: regression coverage
for the gate predicate (missing -> setup forced, installed ->
setup skipped, broken predicate -> non-blocking, unregistered
keys -> behaviour unchanged).
Fixes#22737. Reported by @f-trycua.
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.
- `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.
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.
- `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.
- 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.
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.
- `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.
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.
- `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.
- 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.
