Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/ashishgituser/bunkervm/claude-mdgit clone --depth 1 https://github.com/ashishgituser/bunkervmWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.01578 | $0.01578 |
| Opus 5 | $0.00789 | $0.00789 |
| Sonnet 5 | $0.00316 | $0.00316 |
| Haiku 4.5 | $0.00158 | $0.00158 |
Grade A, and why
bunkervm CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 45 lines — stays where its author put it; the contents beside it link to each section on GitHub.
BunkerVM
Hardware-isolated Linux sandbox for AI agents — Firecracker microVMs with record/replay/snapshot/diff. Pip package bunkervm, console script bunkervm, current version 0.9.4. Core dep is just mcp>=1.0.0.
Repo scope (as of the 2026-08 trim-down, see CHANGELOG.md): trimmed to the core differentiated pitch — time-travel debugging (record/snapshot/restore/diff) — plus MCP support. The desktop GUI app and the dedicated LangChain/OpenAI-Agents-SDK/CrewAI integration modules were untested, high-maintenance, and diluted focus; they were moved to backup/ (gitignored, not tracked, still in git history). Don't recreate them without deliberately deciding to re-scope — check backup/README.md first if something seems to be "missing." SecureAgentRuntime.as_tool() / .as_openai_tool() in agent_runtime.py remain as minimal single-tool adapters (lazy-import the upstream SDK directly, no bundled toolkit class).
Two operating modes (pervasive throughout the codebase)
- Direct mode — talks straight to Firecracker via KVM (Linux only).
Sandbox._vm=bunkervm/vm_manager.py::VMManager, comms over vsock UDS viabunkervm/sandbox_client.py. Supports true Firecracker snapshots (instant, full VM state: memory/fs/processes). - Engine mode — a background daemon (
bunkervm/engine/daemon.py, REST API onlocalhost:9551) owns the only Firecracker access; everything else (CLI, SDK, desktop, MCP server) is a thin HTTP client viabunkervm/engine_client.py/bunkervm/engine/client.py. Used transparently on Windows via a WSL2 bridge (bunkervm/engine/wsl_bridge.py). README lists only Linux (native/dev/kvm) and Windows+WSL2 as supported; macOS is not a documented supported platform (Firecracker requires KVM, which macOS lacks). Cannot do real VM snapshots — some direct-mode-only features raiseRuntimeErrorhere (manualcheckpoint(), true snapshotrestore()). Sandbox.start()tries_resolve_engine()first, then falls back to_start_direct(). Capability parity between the two modes is incomplete by design, not a bug.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 45 lines · 1,578 tokens per session scan A 9b192fd9cd2b
bunkervm CLAUDE.md is an instructions file published in the GitHub repository ashishgituser/bunkervm (0 stars, last pushed 14d ago), licensed MIT. It adds 1,578 tokens to every session, about $0.0079 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
mitos CLAUDE.md
Instructions for mitos-run/mitos, covering claude.md, project overview, operating principles, commands and architecture.
oring CLAUDE.md
Instructions for hexproofdev/oring, covering claude.md, the discipline, hard constraints and standing rejections.
codex-autoresearch AGENTS.md
Instructions for TheGreenCedar/codex-autoresearch, covering agents.md, purpose and scope, canonical sources, stable product boundaries and source and package boundaries.
autoharness AGENTS.md
Instructions for tigerless-labs/autoharness: This repo uses Open Knowledge — collaborative markdown via MCP. .open-knowledge/config.yml (with optional /.open-knowledge/config.yml; CLI/env may override) is the path contract: content.dir is the root for relative paths; content.include lists globs that add markdown…
mirage CLAUDE.md
Claude Code instructions for strukto-ai/mirage, covering claude.md, repo layout, typescript packages, python/typescript parity and module layout.
causal-memory CLAUDE.md
Instructions for JingxuanC/causal-memory, covering causal memory mcp integration, causal memory integration, after acting on a decision and observing the result, when something fails unexpectedly and when a recorded lesson turns out to be wrong.