Skip to content

Ghost in the Droid v1.6.0

Ghost 1.6 replaces the v1 skill runtime with Skill System v2 and ships the Ghost Skill Hub: 117 skills covering 113 Android app packages, each with a score, run records, demo clips and a compatibility record. It also adds typed human-in-the-loop steps, supply-chain checks for community submissions, and a rebuilt skills site. All the evidence in this release was recorded by the maintainers, not independently, and we say so wherever it matters.

  • Skill System v2. Skills are YAML bundles (skill.yaml, SOFT.md, optional helpers.py) with 26 step kinds and 7 validator kinds. ghost skill now has 29 subcommands covering the whole lifecycle.
  • Ghost Skill Hub. ghost skill search | info | install | update | outdated | fork | deprecate, with every downloaded file pinned by sha256 and a version commit. 117 skills across 113 app packages.
  • Scores you can read. Each skill gets a 0 to 100 score from run records, split by evidence class. Downloads and likes earn nothing.
  • Fewer tokens. Median 90% fewer across 99 before/after pairs. The stricter figure on pairs where both runs succeeded is 86%. Details below.
  • Human in the loop. checkpoint pauses for a person (captcha, login, SMS code); ask collects a typed answer.
  • MCP: find_skill first. Four new tools, four private CRM/lead tools removed, still 66 tools in total.
  • Safer submissions. Static analysis, a trust gate, fork-safe emulator runs and seven named supply-chain gates.
  • Opt-in telemetry and guard rails. ghost telemetry on|off|status (off by default), GHOST_DENY_PACKAGES, and a fingerprint-distance check.

Skill system: hard skills replay, soft skills guide

Section titled “Skill system: hard skills replay, soft skills guide”

A skill is skill.yaml plus SOFT.md. skill_class: hard replays without an LLM. skill_class: soft is a guide the agent follows by hand.

  • Fingerprints. Each skill declares OS, OS version range, hardware, app package and app version range. Runs compute a distance of 0 to 5: 0 runs, 1 to 2 warns and runs, 3 or more refuses (--force-fingerprint overrides on the CLI). --strict-fingerprint or GHOST_STRICT_FINGERPRINT=1 requires distance 0. The adapt tier, which would rewrite a skill for a nearby device, is planned for a later release.
  • Popups. Per-skill popup detectors, an OS popup catalog and a safety net for unknown popups.
  • File-bound helpers. python_action: helpers:<fn> loads from the skill’s own helpers.py. Evidence for skills that ship Python must carry a skill_sha256 digest or it earns nothing.
  • Authoring. scaffold, validate, lint, analyze, test, run, probe, doctor, diff, import (Maestro and structured Automate flows) and demo. ghost skill list-registered prints every step and validator kind.
  • Catalog. 116 official skills plus one benchmark skill. New ones include Calculator, Contacts, Clock, Tasks, Markor, Settings, NewPipe, Unitto, Feeder, arXiv search, Open Food Facts and World Clock.
  • Privacy-bounded. Settings 1.1.0 returns only aggregate reads. NewPipe 1.1.0 exposes search metadata only and never opens a video.

The client fetches the registry, downloads each file pinned to its commit, verifies sha256, and installs atomically to ~/.ghost/skills/<author>/<slug>/<version>/. GHOST_SKILLS_DIR moves the install root.

  • Trust labels: official, benchmark, verified, unverified, quarantined. Today 116 skills are official, 1 is benchmark, and none is verified.
  • Quality tiers: draft, runs, proven, hardened, reference (0, 10, 35, 60, 90 points). The registry has 28 proven, 75 runs and 14 draft. None is hardened or reference, because those tiers require independent runs.
  • Site: a /skills/ browser, per-skill pages with demo clips and run history, a leaderboard, a scoring page and a compatibility matrix of every skill against every phone and Android version it was run on. Coverage means tested, not passed.

Token savings: 90% overall, 86% where both runs succeeded

Section titled “Token savings: 90% overall, 86% where both runs succeeded”

We recorded one before/after pair per skill for 99 skills: same prompt, same phone, model claude-haiku-5-5. The median is 90% fewer tokens (10.2x), a median 1.24x speedup and 71% fewer tool calls. Four pairs used more tokens with the skill.

Here’s the caveat. In 28 of the 99 pairs, the run without the skill failed the task, so the 90% figure partly compares a failed manual run with a successful replay. On the 71 pairs where both runs succeeded, the median is 86% fewer tokens and 1.09x faster. It’s still one pair per skill, one model, run by us.

On a six-task native pool (24 paired episodes), 23 passed with the skill and 19 without. That’s exploratory, not statistically significant (p = 0.22).

  • checkpoint pauses a run (captcha, sms, email, login, llm_decision, generic) until a person resumes or aborts it from the dashboard or POST /api/skills/runs/{id}/resume. Default timeout is 600 s.
  • ask collects one typed answer from a human, the calling agent (run_skill_resume or ghost skill resume), or a bounded sub-agent. Paused state is stored privately and traces never keep screenshots.

We tried it on four public captcha demo pages: a fresh agent fills the form, calls a human, and finishes once the person solves it. Clips are sped up, and nothing here bypasses a captcha.

Watch all four captcha demos.

  • New MCP tools: find_skill, run_skill, run_skill_resume, get_skill_notes. The server tells the agent to call find_skill before driving an app by hand.
  • Profiles: GHOST_MCP_PROFILE selects skills-first (17 tools), soft-only (15) or hand-only (10). An unknown or empty value serves all 66. GHOST_MCP_HIDE_TOOLS=a,b drops named tools.
  • ghost skill is now the v2 CLI. The legacy parser stays reachable as android-agent skill ....
  • Submission: ghost skill submit <dir> [--push] and ghost skill verify <ref> [--push] open PRs; without --push they print the plan.
  • Device scoping: GHOST_DEVICE (or ANDROID_SERIAL) limits discovery to one connected device.
  • PR jobs run on ephemeral hosted runners with no secrets and a read-only token, using pull_request and never pull_request_target. Signing and registry commits happen only after a maintainer merge.
  • Seven gates: network-exfil, aliased-exec, soft-injection, typosquat-slug, unbound-helper, wrong-app and stale-app. A frozen local corpus of 22 cases (20 expected catches, 2 benign controls) passes. That’s fixture coverage, not a general security guarantee.
  • Trust gate: emulator runs start only for allowlisted apps with no submitter Python, or after a maintainer approves that exact commit.
  • Run records are signed with SSH Ed25519 keys. A signature shows a record wasn’t altered, not where it came from.
  • GHOST_DENY_PACKAGES=pkg1,pkg2 makes tool calls refuse named apps. It’s an accident guard inside one process, not a sandbox.
  • Telemetry is opt-in (default off, DO_NOT_TRACK=1 wins). The collection endpoint is staged, not deployed.
  1. Four MCP tools removed: lookup_lead, list_unread_leads, crm_lookup_contact, crm_list_unread_messages. No replacement; they depended on a private lead database.
  2. ghost skill install takes hub refs only (author/slug[@version]). For GitHub-URL, local-path or bare-name installs, use android-agent skill install ....
  3. v1 skills: run ghost skill migrate <path> [--author HANDLE] [--dry-run]. It keeps every v1 file, so the v1 runtime keeps working.
  4. Bundled skills moved. The five v1 skills under gitd/skills/ (TikTok, TikTok iOS, Safari, Play Store and an email-sending skill) are gone. Use the hub copies under skillhub/ghost/ or an installed copy.
  5. list_devices output changed to <id> model=<model> rows. Update anything parsing serial (model), and pass only the first column as device.
  6. Fingerprint distance 3 or more is now refused. There’s no force flag on the MCP tool.
  7. No config or environment variables were renamed. New ones are additive.
Terminal window
pipx upgrade ghost-in-the-droid # or: pip install -U ghost-in-the-droid
ghost skill search calculator
ghost skill info ghost/calculator
ghost skill install ghost/calculator
ghost skill run ghost/calculator calculate --param expression='1234*5' --device <device-id>
ghost skill migrate <path-to-v1-skill> --dry-run # then again without --dry-run

MCP clients: restart the server after upgrading, and optionally set GHOST_MCP_PROFILE=skills-first.

  • Every recorded run is maintainer-run. There are zero independent runs, so no skill can reach Hardened or Reference, and the device grid doesn’t establish cross-device reliability.
  • Proven means the author’s own passing runs plus a demo and validators, not independent proof.
  • The token and speed medians come from one pair per skill with one model.
  • Planned, not in 1.6.0: ghost skill mine, ghost skill try and the adapt tier, which is planned for a later release.
  • The auto-login state runner is beta.
  • A blank cell in the compatibility matrix means not tested, not broken.

Read the 1.6 blog post for the story behind the numbers, or the full GitHub release notes for every flag, fix and CI change.

Join the Ghost Discord to share what you build, report failures and get support. Help teach the ghost.