A Rust command centre for autonomous security operations.

Most "agentic" tooling is a confidence trick performed with logs. Serpantoxide takes the less glamorous view that orchestration should be explicit, typed, inspectable, and difficult to romanticise. It is a Rust runtime for running a crew of security workers with a proper operator surface, shared state, browser control, topology intelligence, and just enough suspicion about its own abstractions to remain useful.

In plain English: this is a working offensive-security console. It plans, delegates, watches, revises, records, and reports. It has a production TUI, an experimental macOS GPUI shell, and very little interest in pretending that "AI" is a substitute for instrumentation.

Because the usual arrangement is intolerable.

• Python prototypes become accidental constitutions.
• Agent frameworks multiply concepts the way damp basements multiply fungus.
• Tool calls vanish into stringly typed fog.
• Interfaces soothe the operator precisely when they ought to confess uncertainty.

Serpantoxide is the opposite design instinct: fewer illusions, harder edges, better visibility.

• A live terminal UI with telemetry, worker logs, topology views, inspection panes, and report generation.
• An experimental native macOS shell built on the same runtime.
• A mission-aware crew orchestrator that can infer intent, publish a checklist, spawn, monitor, and synthesise multiple workers.
• Worker agents that operate as iterative tool-calling loops rather than decorative one-shot prompts.
• Native tools for terminal, browser, web_search, notes, nmap, sqlmap, osint, hosting, image_gen, and evm_chain.
• Persistent findings in loot/notes.json, filtered by active target when a scope is set.
• Prompt history, multiline completion lists, target-aware autocomplete, shared operator note storage, pause-and-correct control, and persisted mission/config state.
• A lightweight graph model that turns findings into something closer to intelligence.
• Deterministic mock mode when provider keys are absent, because pretending otherwise would be vulgar.

Serpantoxide is arranged around four deliberate layers:

1. main.rs boots the selected frontend.
2. runtime.rs and automation.rs provide the shared command, event, snapshot, and operator-control model.
3. orchestrator.rs thinks in campaigns.
4. worker_agent.rs does the grubby work and returns with evidence.

That separation matters. The orchestrator is there to think at a level above a shell command. The workers are there to prevent those thoughts from floating away into rhetoric.

• Rust toolchain with cargo
• Chromium or a compatible browser runtime for chromiumoxide
• Optional native binaries:
  • nmap
  • sqlmap
  • holehe
  • sherlock
  • theHarvester
• Optional environment variables:
  • OPENROUTER_API_KEY
  • DEEPSEEK_API_KEY
  • TAVILY_API_KEY
  • GOOGLE_API_KEY
  • ETHERSCAN_API_KEY
  • EVM_RPC_URL
  • LLM_MODEL

cargo run

If startup goes sideways before the UI appears, inspect:

tail -n 80 /tmp/serpantoxide-startup.log

# Default TUI
cargo run

# Experimental macOS shell
cargo run -- --gpui

# Force the TUI explicitly
cargo run -- --tui

scripts/package_macos_app.sh
scripts/package_macos_app.sh --target x86_64-apple-darwin --zip

/agent <task>        Run a focused autonomous assessment
/crew <task>         Run multi-agent crew mode
/preset [name]       Show or select a mission preset
/presets             List mission presets
/target <host>       Set the active target
/tools               Show worker capabilities
/notes [category]    Show stored findings
/store <cat> <text>  Store an operator note in shared knowledge
/config              Show runtime config
/config set ...      Update runtime config
/pause [all|crew|agent-id]
/resume [all|crew|agent-id] [instruction]
/memory              Show graph-derived intelligence
/topology            Open the interactive topology explorer
/prompt              Show the crew prompt
/report              Generate a markdown report
/models              Open the model picker
/keys                Open the provider key editor
/clear               Clear telemetry
/help                Show help
/modes               Show mode and prefix help
/quit                Exit

The TUI prompt also supports:

• freeform non-slash input as an implicit crew mission
• multiline paste
• Up and Down prompt history navigation when no completion list is open
• current-line completion inside multiline drafts with a scrollable suggestion list
• Up and Down suggestion navigation plus Tab accept for completions
• target-aware LLM ghost-text completion for single-line prompts
• runtime lane enforcement that blocks off-scope tool calls and forces a replan
• Esc to pause all active automation, inject corrective instructions, and resume
• P to pause the selected worker and inject a worker-specific correction

Workers can call:

• terminal
• browser
• web_search
• notes
• nmap
• sqlmap
• osint
• hosting
• image_gen
• evm_chain

And if subtlety is getting in the way, forced intent prefixes are available:

NMAP: <host>
SQLMAP: <url>
BROWSER: <url>
SEARCH: <query>
TERMINAL: <command>
OSINT: <tool and target>
HOSTING: <action and path>
IMAGE: <prompt>
EVM: <action and address/query>

Serpantoxide stores local runtime state in .serpantoxide_config. If OPENROUTER_API_KEY or DEEPSEEK_API_KEY is present, it loads those provider catalogs into /model. If both are absent, the runtime drops into deterministic mock behaviour. This is not a scam. It is simply the difference between a live provider and a rehearsal.

Current persisted fields:

{
  "selected_model": "openai/gpt-4o",
  "selected_preset": "auto",
  "last_target": "example.org",
  "max_iterations": 16
}

You can inspect and update the runtime config from inside the app:

/config
/config set max_iterations 24

Typical configuration:

OPENROUTER_API_KEY=...
DEEPSEEK_API_KEY=...
TAVILY_API_KEY=...
GOOGLE_API_KEY=...
ETHERSCAN_API_KEY=...
EVM_RPC_URL=...
LLM_MODEL=openai/gpt-4o

src/
  main.rs
  automation.rs
  runtime.rs
  startup_trace.rs
  tui.rs
  gpui_app.rs
  orchestrator.rs
  pool.rs
  worker_agent.rs
  mission.rs
  llm.rs
  browser.rs
  notes.rs
  graph.rs
  terminal.rs
  nmap.rs
  sqlmap.rs
  osint.rs
  hosting.rs
  image_gen.rs
  evm_chain.rs
  prompts.rs
  events.rs
  config.rs

• docs/OPERATOR_RUNTIME_SPEC.md
• docs/OPERATIONS.md
• docs/TOOL_REFERENCE.md
• docs/DEVELOPMENT.md
• docs/GIT_SPLIT.md

cargo fmt
cargo check
cargo test

Use this only against systems you are authorised to assess. Good tooling does not suspend ethics. It merely removes the excuse of incompetence.