Skills/Glama/ghost-hands

ghost-hands@1.11.0

CAUTION· not started

Publisher claims: "Ghost Hands gives agents governed, recorded, replayable hands on the web through a zero-dependency Python CDP stack, classifying every action before it runs and writing it to a provenance trail." Observed in sandbox: no network, touches nothing outside the project, writes no memory.

Registry GlamaPublisher littlestjames82-sys · publishing for 8 daysInstalls 0Last detonated 2 h ago
What would fix thisWatchDispute verdict

Coverage

Not started: dependency install failed; the server itself was not observed.
Tested
the dependency install step · every file read, file written and host contacted by what ran
Not tested
server startup · the tool list · tool behavior

What this means for the label: Caution here covers what ran. The server's tools were not exercised with a real credential, so nothing about their behavior is asserted; the credential summary below says what to do with that.

Before you give it your token

It will ask for
ANTHROPIC_API_KEYGHOST_HANDS_ANDROID_TOKENGHOST_HANDS_ANTHROPIC_API_KEYGHOST_HANDS_API_KEYGHOST_HANDS_FALLBACK_API_KEYGHOST_HANDS_SERVE_TOKENGHOST_HANDS_WEBHOOK_URL or --max-llm-tokens / --android-tokenConfiguration, not credentials: ANTHROPIC_BASE_URL, ANTHROPIC_MODEL, GHOST_HANDS_ANDROID_BRIDGE, GHOST_HANDS_ANTHROPIC_BASE_URL, GHOST_HANDS_ANTHROPIC_MODEL, GHOST_HANDS_APPROVAL_CHANNEL, GHOST_HANDS_APPROVER, GHOST_HANDS_BASE_URL, GHOST_HANDS_BATCH, GHOST_HANDS_CACHE, GHOST_HANDS_CHROME, GHOST_HANDS_DECIDER, GHOST_HANDS_DRIVER, GHOST_HANDS_FALLBACK_BASE_URL, GHOST_HANDS_FALLBACK_MODEL, GHOST_HANDS_FIELD_VALUES, GHOST_HANDS_GHOSTGUARD_AUDIT, GHOST_HANDS_GHOSTGUARD_BRIDGE, GHOST_HANDS_GHOSTGUARD_NODE, GHOST_HANDS_GHOSTGUARD_POLICY, GHOST_HANDS_JSON_MODE, GHOST_HANDS_MODEL, GHOST_HANDS_POLICY, GHOST_HANDS_POSTCONDITIONS, GHOST_HANDS_QUIET, GHOST_HANDS_RETRIES, GHOST_HANDS_RETRY_BACKOFF, GHOST_HANDS_START_URL, GHOST_HANDS_VALIDATE, GHOSTBUS_HOME, GHOSTBUS_KEY, GHOSTBUS_URL, GHOSTBUS_WORKSPACE
Where it talks to
no vendor host named in the codeAlso demo.local, shop.local, app.local, toggle.local, noop.local, x.local, evil.local, other.local, bank.local, heal.local, proxy.local, form.local, conf.local, trap.bench, v112.local, v113.local, v114.local, v115.local, v116.local, eval.local, openrouter.ai — not the publisher's domain; check what for before connecting.
What the honeypot saw
no credential files read · no network
What we could not test
server startup · the tool list · tool behavior

Connect it with a scoped credential, under the guard, and read the third-party hosts first. The code names demo.local, shop.local, app.local, toggle.local, noop.local, x.local, evil.local, other.local, bank.local, heal.local, proxy.local, form.local, conf.local, trap.bench, v112.local, v113.local, v114.local, v115.local, v116.local, eval.local, openrouter.ai as well as its own vendor. The guard injects ANTHROPIC_API_KEY, GHOST_HANDS_ANDROID_TOKEN, GHOST_HANDS_ANTHROPIC_API_KEY, GHOST_HANDS_API_KEY, GHOST_HANDS_FALLBACK_API_KEY, GHOST_HANDS_SERVE_TOKEN, GHOST_HANDS_WEBHOOK_URL only into requests to the host you bind (and its subdomains) and refuses the rest.

$vet-skills run --net declared -- <the server's command>

Never paste a credential into a chat, a ticket or a report, including ours. How the guard holds it.

Is [email protected] safe to install?

KillerSkills ran it in the KillerSkills honeypot sandbox (2 scenarios; not started: dependency install failed; the server itself was not observed) on 2026-11-14 and scored it 73/100 (Caution). It is not known to be malicious, but it has behaviour or code a security policy should decide on before install. Notable findings: URL with embedded credentials; Credential pass-through configuration; Credential pass-through configuration. Summary: no network, touches nothing outside the project, writes no memory.

Behavior trace

Honeypot run · 0 s · 2 scenarios · 0 deviations
t+CategoryObserved action
0.0sprocessno install scripts in bundle
0.2sprocessdependency install exit 1 ·
0.2sclaimednot started: dependency install failed (exit 1); the server itself was not observed
0.2sprocessscript backend runs setup only
0.2sprocessno install scripts to re-run

Static findings

MEDIUM
File type combined with upload actionsrc/ghost_hands/busagent.py:622 · File type combined with upload action — agent may exfiltrate local data filesself.log(f"complete failed ({exc}); posting the report as a comment")Fix: Remove the call, or declare the host in frontmatter and explain in the README what is sent and why.
AST07
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/doctor.py:460 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
MEDIUM
File type combined with upload actionsrc/ghost_hands/busagent.py:602 · File type combined with upload action — agent may exfiltrate local data filesself.log(f"trail upload failed: {exc}")Fix: Remove the call, or declare the host in frontmatter and explain in the README what is sent and why.
AST07
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/bench.py:1128 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/doctor.py:116 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
HIGH
URL with embedded credentialssrc/ghost_hands/bench.py:949 · URL with embedded credentials — may be phishing or credential leakingd = ChromiumDriver(proxy="http://user:[email protected]:3128")Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
AST07
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/drivers.py:1492 · Python subprocess execution detectedproc = subprocess.Popen(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/busagent.py:731 · Python subprocess execution detectedproc = subprocess.Popen(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
HIGH
Credential pass-through configurationsrc/ghost_hands/phone_approver.py:247 · Credential pass-through configuration — agent relays sensitive credentials"no Ghost Hands pairing token: pass token= or set "Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
AST02
MEDIUM
MCP server using network transport (SSE/HTTP)src/ghost_hands/bench.py:51 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom .mcp_server import HandsSession, handle_requestFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04
HIGH
Credential pass-through configurationsrc/ghost_hands/drivers.py:17 · Credential pass-through configuration — agent relays sensitive credentialsproxy credentials in ``--proxy-server``), and stale-target detection inFix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
AST02
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/doctor.py:484 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
HIGH
Credential pass-through configurationsrc/ghost_hands/android_driver.py:171 · Credential pass-through configuration — agent relays sensitive credentials"no Ghost Hands pairing token: pass token= or set "Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
AST02
LOW
Prompt injection attemptANNOUNCE_READY.md:23 · Prompt injection attempt — 'ignore previous instructions' · informational: not an instruction file- The seatbelt policy pack: standard posture plus deny patterns for injection-shaped target text ("ignore previous instructions" and friends), with tests guarding benign lookalikes.Fix: Keep SKILL.md to visible, descriptive text. Remove comments and encoded text that address the agent.
AST01
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/ghostguard.py:278 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
LOW
URL to a private-network addresstests/test_doctor.py:166 · URL to a private-network address — a LAN or development targetenv = bridge_env("http://10.255.255.1:8378")Fix: Replace encoded or assembled strings with plain literals and remove undeclared network calls.
AST07
MEDIUM
Process kill commandsrc/ghost_hands/mcp_server.py:753 · Process kill command — agent can terminate processes# A hostile or malformed message must never kill theFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.
AST05
MEDIUM
Python subprocess execution detectedtests/test_busagent.py:348 · Python subprocess execution detectedproc = subprocess.Popen(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
MEDIUM
Python subprocess execution detectedtests/test_export_v2.py:46 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
MEDIUM
MCP server using network transport (SSE/HTTP)tests/test_policies.py:211 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom ghost_hands.mcp_server import HandsSessionFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04
MEDIUM
MCP server using network transport (SSE/HTTP)tests/test_policies.py:219 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom ghost_hands.mcp_server import HandsSession, handle_requestFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04
LOW
Refusal suppressionCHANGELOG.md:187 · Refusal suppression — attempting to prevent safety responses · informational: not an instruction filerunner-internal heals are never refused). Kinds with noFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.
AST01
MEDIUM
Process kill commandsrc/ghost_hands/drivers.py:654 · Process kill command — agent can terminate processespass # a broken handler must never kill the transportFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.
AST05
MEDIUM
MCP server using network transport (SSE/HTTP)tests/test_policies.py:201 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom ghost_hands.mcp_server import HandsSessionFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04
MEDIUM
MCP server using network transport (SSE/HTTP)tests/test_chromium_v3.py:21 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom ghost_hands.mcp_server import HandsSession, handle_requestFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04
LOW
Prompt injection attemptCHANGELOG.md:696 · Prompt injection attempt — 'ignore previous instructions' · informational: not an instruction file"ignore previous instructions", "disregard your instructions",Fix: Keep SKILL.md to visible, descriptive text. Remove comments and encoded text that address the agent.
AST01
MEDIUM
MCP server using network transport (SSE/HTTP)tests/test_mcp_parity.py:15 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom ghost_hands.mcp_server import HandsSession, handle_requestFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04
MEDIUM
Process kill commandsrc/ghost_hands/mcp_server.py:310 · Process kill command — agent can terminate processesthe dispatcher and kill the server (v0.9.1)."""Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.
AST05
LOW
Prompt injection attemptREADME.md:212 · Prompt injection attempt — 'ignore previous instructions' · informational: not an instruction file(`policies.INJECTION_PATTERNS`: "ignore previous instructions",Fix: Keep SKILL.md to visible, descriptive text. Remove comments and encoded text that address the agent.
AST01
MEDIUM
Python subprocess execution detectedsrc/ghost_hands/phone_proof.py:143 · Python subprocess execution detectedproc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
AST05
MEDIUM
MCP server using network transport (SSE/HTTP)tests/test_mcp.py:4 · MCP server using network transport (SSE/HTTP) — data sent over networkfrom ghost_hands.mcp_server import HandsSession, handle_requestFix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.
AST04

Cost to the agent

Instructions loaded
≈40 tok
1 file · 81 in bundle
Tool list per prompt
≈0 tok
0 tools · 0 per tool
Handshake
no reply
launch to tools/list, stdio
Undocumented tools
0
description under 20 chars · avg 0

Tokens estimated at four characters each. The tool list is re-read by the agent on every turn, so it is the number to compare between servers that do the same job: ranked across the index.

Known vulnerabilities

Vulnerability advisories (OSV) cover npm and PyPI packages; this registry has no advisory database.

What would make this Verified

7 points to 80 · weights: behavior 45%, static 25%, prompt 15%, publisher 15%
  1. Unclaimed publisher≈ +8 on the totalClaim littlestjames82-sys: a verified publisher gets +50 on publisher trust. It is free and usually instant with GitHub sign-in.
  2. URL with embedded credentials≈ +4 on the totalStop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
  3. Credential pass-through configuration≈ +4 on the totalStop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
  4. New publisher≈ +3 on the totalTime clears this on its own after 14 and 30 days. Claiming the handle (+50) is the fastest way past it.
  5. File type combined with upload action≈ +2 on the totalRemove the call, or declare the host in frontmatter and explain in the README what is sent and why.
  6. Python subprocess execution detected≈ +2 on the totalBuild commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
Publish a new version to Glama, then queue it here. Not your skill? Claim the handle to get verdict changes by email.Claim littlestjames82-sysPropose a patch to the maintainer