PlamenTSV/plamen
Plamen is an open-source Python orchestrator that drives 18-100 Claude Code or OpenAI Codex CLI agents across 40+ phases to autonomously audit smart contracts (EVM/Solidity, Solana/Anchor, Aptos Move, Sui Move, Soroban, DAML/Canton) and L1 node-client infrastructure (Go/Rust consensus, p2p, mempool code), producing a severity-tiered AUDIT_REPORT.md with verified proof-of-concept exploits for each finding. It integrates MCP servers for Slither, ChromaDB and Solodit, ingests vulnerability corpora from Solodit, DeFiHackLabs and Immunefi to ground findings, and runs a PTY-supervised worker-pool architecture with checkpoint/resume, artifact-based completion gating, and a dedicated L1 mode with consensus-invariant and network-surface depth agents.
What it is
Plamen's installation and operation surface — the concrete setup workflow, backend trade-offs, and the ordering/config constraints that determine whether an audit run actually works. The summary tells you what Plamen does; this tells you what you have to get right to run it.
How it works
Two backend CLIs are supported and you install at least one. Claude Code has the broadest MCP support (Slither, ChromaDB, Solodit). OpenAI Codex CLI works but doesn't yet support every MCP server — pure-LLM phases fall back to WebSearch. You can install both and pick per-run in the audit wizard.
plamen install is non-interactive (symlinks, config, Python deps, hook self-heal) and safe in any context including CI and headless servers. plamen setup runs the install then drops into an interactive toolchain wizard (Foundry, Solana CLI, Anchor) — run it from a real terminal; in a non-TTY context it exits cleanly after install rather than crashing on the picker.
Preconditions / Affected
Python 3.11–3.12 + pip, Node.js 18+, Git. macOS needs xcode-select --install; Windows needs Developer Mode on (for symlinks). The optional RAG vulnerability DB needs ~6GB RAM.
What to do
Clone with submodules — ZIP downloads silently omit custom-mcp/slither-mcp/, custom-mcp/farofino-mcp/, and the three opengrep-rules/* sets:
git clone --recurse-submodules https://github.com/PlamenTSV/plamen.git ~/.plamen
cd ~/.plamen && python3 plamen.py installIf you already cloned without submodules: git submodule update --init --recursive from inside ~/.plamen/ before install.
To drive setup from an AI assistant, paste SETUP.md only — it has error handling and blocks the heavy RAG build from a non-TTY context. Do not paste docs/setup.md or docs/getting-started.md; those are human manuals with the RAG build inline.
Put the Solodit key where subprocesses can see it — a shell export is not enough, because Claude Code/Codex spawn non-interactive subshells that don't source .bashrc/.zshrc. Add SOLODIT_API_KEY to ~/.claude/settings.json → "env" (or ~/.codex/config.toml → [env]) before running plamen rag. Free key from solodit.cyfrin.io.
Verify without spending API calls: plamen doctor. Resume an interrupted audit: plamen resume.
Caveats
On externally-managed Python (Homebrew, Ubuntu 23.04+), plamen install auto-adds --break-system-packages and writes into system site-packages, printing a stderr notice. To isolate deps instead, activate a virtualenv first and set PIP_BREAK_SYSTEM_PACKAGES=0. Codex runs lose MCP-backed detection on unsupported phases, degrading to WebSearch.
Technologies named here
- Plamen
- EVM/Solidity smart contracts
- Solana/Anchor
- Claude Code CLI
- OpenAI Codex CLI
- Aptos Move
- Slither
- Sui Move
- ChromaDB
- Soroban/Stellar
- DAML/Canton
- Solodit
- Foundry
- L1 node-client infrastructure
- Solana CLI / Anchor
- Aderyn
- DeFiHackLabs exploit corpus
- Immunefi
- unified-vuln-db
- solana-fender
- opengrep-rules / aptos-move-rules / decurity-rules
- MEDUSA
- Trident
- scip-go / rust-analyzer
- Claude Code
- Codex CLI
- Model Context Protocol
- Solidity
- Python
- Trail of Bits
Techniques named here
- Verified PoC exploit generation
- Force-by-default PoC gate
- M2 multi-axis coverage meta-pass
- Mechanical recall gates
- Cross-OS hygiene gate
- Consensus safety / fork choice analysis
- P2P DoS/eclipse attack analysis
- Mempool asymmetric DoS analysis
- BLS aggregation and light client proof analysis
- State sync/pruning and execution client hardening analysis
- Concurrency safety and unsafe-block auditing
- Static analysis via Slither
Related in AppSec
Flowise MCP Security Bypass Enables RCE
Inside an Oracle Database SQL Injection Attack | Huntress
OASIS
