stratum-mcp@0.2.114
CAUTION· not startedPublisher claims: "Stratum MCP server for Claude Code — structured execution, typed contracts, postcondition enforcement" Observed in sandbox: no network, touches nothing outside the project, writes no memory.
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
STRATUM_GUARD_OVERRIDE_TOKENConfiguration, not credentials: CLAUDE_MODEL, STRATUM_CODEX_ALLOW_WRITE, STRATUM_CODEX_STREAM_LIMIT_BYTES, STRATUM_DEFAULT_AGENT, STRATUM_GUARD_ALLOW_COMMANDS, STRATUM_GUARD_CMD_TIMEOUT_S- Where it talks to
- no vendor host named in the code
- 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. It asks for STRATUM_GUARD_OVERRIDE_TOKEN and talks to its vendor. Give the credential to the guard, not the server: the value never enters the server's environment and only travels to the host you bind. What the tools do with a real credential was not observed.
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 2026-11-19 and scored it 77/100 (Caution). It is not known to be malicious, but it has behaviour or code a security policy should decide on before install. Notable findings: Instructs the agent to modify its persistent memory or instruction files; Instructs the agent to modify its persistent memory or instruction files; Instructs the agent to modify its persistent memory or instruction files. Summary: no network, touches nothing outside the project, writes no memory.
Behavior trace
Honeypot run · 1 s · 2 scenarios · 0 deviationsStatic findings
return eval(code, {"__builtins__": {}, **_ENSURE_BUILTINS}, {bind: value})Fix: Put hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.diff_result = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.return bool(eval(Fix: Put hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.…rns. **After `stratum_audit`:** The `write_memory` step produces the entries — write them to `.claude/memory/MEMORY.md` (create directory and file if missing). Tag each entry with the skill it appli…Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.**After the fix:** Append to `MEMORY.md`:Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.5. Write conclusions to `MEMORY.md` at `.claude/memory/MEMORY.md` (create if missing)Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.**After migration:** Append to `MEMORY.md`:Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.…ule with known fragility, a class of issue this project tends to have)? If yes, append a one-liner to `MEMORY.md`: ``` [stratum-review] <pattern> — e.g. "auth middleware bypasses rate limitin…Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.--- name: stratum-onboard description: Read a codebase cold and write a project-specific MEMORY.md from scratch — architecture patterns, test conventions, module constraints, kno…Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.result = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.value = eval(code, {"__builtins__": {}, **_ENSURE_BUILTINS}, {"result": result})Fix: Put hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.result = subprocess.run(cmd, capture_output=True, check=False)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.untracked_result = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.return subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.**After Phase 4 `stratum_audit`:** For each step with `attempts > 1`, ask: does the retry reveal a project-specific pattern worth capturing? Append one-liners to `MEMORY.md`:Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.**After `stratum_audit`:** If any step needed retries, append to `MEMORY.md`:Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.return bool(eval(code, {"__builtins__": {}, "None": None, "True": True, "False": False,Fix: Put hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.**After each extraction:** If a step needed retries because of an unexpected dependency or import constraint, append to `MEMORY.md`:Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.…ious coupling, a test convention, an invariant that must be preserved)? If yes, append a one-liner to `MEMORY.md`: ``` [stratum-feature] tests use factory_boy fixtures, not pytest fixtures — …Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.result = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.out = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.result = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.res = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.return bool(eval(Fix: Put hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.ev.parse_predicate_statement("os_system('rm -rf /')")Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# v0.2: terminal status — set to "killed" when gate kill fires with null on_killFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# Stratum hook (T2-M3): append a timestamped summary note to MEMORY.md atFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.subprocess.run(["git", "commit", "-qm", "init"], cwd=path, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.# Graceful shutdown of THIS controller: kill only the writers it owns.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", *args], cwd=tmp_path, capture_output=True, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.subprocess.run(["git", "add", filename], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out."""True when the kill switch env var is set to a truthy value."""Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.- kill maps to killFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# cancel of a reparentable task, do NOT interrupt (would killFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.prompt="status kill test",Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# detach-don't-kill (review #1)Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# run with no ownership stamp, never kill it while its child is alive.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "init", "-q", "-b", "main"], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.# Timeout — hard kill the subprocessFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# After kill the goal state and flow may be cleaned up; if decision_gates isFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "add", "README"], cwd=path, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out."""Run a regex search in a subprocess with a hard kill timeout.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "init", "-q"], cwd=path, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.process-group interrupt, and the "finally must not kill the detached child"Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "config", "user.name", "Test"], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out."""kill → status surface shows decision='kill', outcome='kill'."""Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "add", ".gitignore"], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.# T2-F5-RESUME (review #1): detach-don't-kill — bypass the WHOLEFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "commit", "-q", "-m", "ignore"], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out."kill routes to on_kill target (or terminates the flow if null). "Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.| `/stratum-onboard` | Read a new codebase cold and write project-specific `MEMORY.md` |Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct."""A successful completion under require=all does NOT cascade-kill siblings."""Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "config", "user.name", "t"], cwd=path, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.| `/stratum-onboard` | Read a new codebase cold and write project-specific `MEMORY.md` |Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.# and kill the sibling reparented children so their readers terminalize.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# Neutral wording covers install, update, and re-chmod casesFix: Request the narrowest permission that works; replace wildcards with the exact paths or hosts.# interrupt (killpg) + finally-no-kill teardown contractFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.proc = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.subprocess.run(["git", "config", "user.email", "t@t"], cwd=path, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out."as an explicit kill outcome. "Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# closing the generator (shutdown/cancel) must NOT kill the detached childFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.still reparentable. The ``finally`` does NOT kill the child — only anFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.# shutdown — leave it `running` and don't kill the durable child.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell."Finding 2a: confirm→approve, reject→revise, kill→kill mapping required."Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.subprocess.run(["git", "config", "user.email", "[email protected]"], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.subprocess.run(["git", "commit", "-q", "-m", "init"], cwd=repo, check=True)Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.sha = subprocess.run(Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.# Content matches but execute bit dropped — re-chmod onlyFix: Request the narrowest permission that works; replace wildcards with the exact paths or hosts."fires an auto-kill with resolved_by: system following the same on_kill routing "Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.dead = subprocess.Popen(["true"])Fix: Build commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out."killed" — flow terminated (kill with null on_kill)Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell."""D3: kill → outcome=kill."""Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.ev.parse_predicate_statement("os_system('rm -rf /')")Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.Uses multiprocessing to enforce hard kill on stuck regex workers.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.removed.append("CLAUDE.md")Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.changed.append("CLAUDE.md")Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.changed.append("CLAUDE.md")Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct."""Write .claude/mcp.json and append Stratum block to CLAUDE.md."""Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.Cost to the agent
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
No advisories in OSV for stratum-mcp@0.2.114 as of 4 h ago. This is the package itself, not its dependencies.
What would make this Verified
3 points to 80 · weights: behavior 45%, static 25%, prompt 15%, publisher 15%- Text that tells the agent to ignore its instructions≈ +9 on the totalDelete phrases like "ignore previous instructions", "override", "bypass" from SKILL.md and tool descriptions. Describe what the skill does; do not instruct the agent about other rules.
- Unclaimed publisher≈ +8 on the totalClaim smartmemory: a verified publisher gets +50 on publisher trust. It is free and usually instant with GitHub sign-in.
- Python exec or eval of computed code≈ +2 on the totalPut hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.
- Python subprocess execution detected≈ +2 on the totalBuild commands from fixed arguments, never from interpolated strings; prefer an API client over shelling out.
- Destructive file system command detected≈ +2 on the totalBuild commands from fixed arguments; do not interpolate untrusted strings into a shell.
- Process kill command≈ +2 on the totalBuild commands from fixed arguments; do not interpolate untrusted strings into a shell.