Skip to content

Use Bravos from an agent

Bravos does its work through the coding agent in your session. bravos setup installs the same 16 skills for the agents you choose, hooks that brief the agent at session start and coach it while you work, and the MCP servers it calls. This page covers what each agent gets and how to drive it.

Agent Skills Session start You see
Claude Code .claude/skills SessionStart hook in .claude/settings.json; MCP in .mcp.json the banner, at launch
Codex .agents/skills hook in .codex/hooks.json (trust it once in /hooks); MCP in .codex/config.toml the Bravos card, shown as a hook message with your first message
Cursor .agents/skills, through .cursor/rules/bravos.mdc sessionStart hook in .cursor/hooks.json the agent’s first reply opens with the card
opencode .claude/skills and .agents/skills, through /bravos… commands in .opencode/commands a plugin in .opencode/plugins adds the briefing to the system prompt; guardlink MCP in opencode.json the agent’s first reply opens with the card
omp the Claude Code skills and MCP .omp/extensions/bravos.js the banner, above the editor until your first prompt
Antigravity .agents/skills a SessionStart group in .agents/hooks.json the briefing goes to the model only

--agent takes claude, codex, cursor, agy, opencode, omp or all. Only the agents you name get their files. Without --agent, setup detects them from the repository; a .cursor/ folder adds Cursor’s rule and hook. VS Code’s Copilot agent reads hooks from .claude/settings.json as a preview, so the briefing can load there too. Gemini CLI uses the published plugin, not bravos setup.

bravos guide harnesses prints the same, in the installed version’s words. Cursor’s hook keeps any hooks your team already has in .cursor/hooks.json, and bravos uninstall removes only Bravos’s entry.

Server Launch When it is registered
guardlink guardlink mcp always
cxg cxg mcp when cxg is installed
codegraph bravos graph mcp when the code graph bundle is staged (Team)

bravos must be on PATH where the agent runs for the code graph server to start. After it is added, reconnect MCP in your agent once (/mcp in Claude Code). Entries you wrote yourself are kept.

At session start a hook gathers the repository’s security picture: open exposures, confirmed exploits, what to test next, stale claims, tickets, teammates’ open pull requests and the recorded intent. It hands that to the model, so no command is needed. On a fresh start you also see the banner or card in the table above; the model does not see it.

The briefing names the plan (Plan:) and where annotations go (Annotations:). It also says when the repository’s Bravos files are behind the installed version; the agent then runs bravos setup itself, which only brings what is installed up to date. Installing a new Bravos, GuardLink or cxg waits for you. Share Bravos with your team shows a real briefing.

While you work, three more hooks run in Claude Code and Codex:

  • The coach. Before each prompt it may add one line for the agent: a question about issues or threats builds the dashboard while the agent answers, and from the second prompt it reminds once per topic about claims waiting to be confirmed, coverage, or an old dashboard. BRAVOS_COACH=off turns it off for a run, and bravos prefs set coach off --by "<your name>" for the clone.
  • Threat cards. When an edit adds a critical or high @exposes that no test has confirmed or refuted, a card appears in your terminal with one next step. The agent never confirms, refutes or accepts a threat itself. BRAVOS_THREAT_CARDS=off or bravos prefs set cards off turns cards off.
  • Background annotation. While coverage is under 100%, each new session asks the agent to start one annotation pass in a background subagent beside what you asked. bravos prefs set annotate off --by "<your name>" switches it off for the clone.

Before the first commit of a session, the skills ask whether to commit or leave the files for review. bravos prefs set commits commit|leave|ask --by "<your name>" records your answer for the clone.

Ask for what you want (“is this PR safe to open?”) and the agent picks the skill, or name it:

Agent Syntax
Claude Code, opencode /bravos-ship
Codex $bravos-ship
omp /skill:bravos-ship
elsewhere “use the bravos-ship skill”

/bravos is the starting point: it reads the briefing and names one next step, with the reason.

Skill What it does Plan
bravos reads the briefing, explains the loop, routes to the rest Free
bravos-context learns what the app intends from your design documents and the code, and writes .guardlink/intent.md. See context Free
bravos-annotate extends the threat model in a budgeted slice or deep rounds, with the lint gate Free
bravos-threat-hunt “identify all the threats”: maps the attack surface, runs SAST, records every finding, then reports from what was recorded Free; uses the code graph and SAST on Team
bravos-templatize turns open exposures into validated cxg templates and retest kits Free
bravos-env brings up a target you own: loopback by default, reviewed scope.yaml, your typed authorization Free
bravos-prove runs the authorized test and records outcomes through guardlink’s ledger Free
bravos-recheck after a fix: re-verifies stale claims and replays retest kits Free
bravos-resync carries annotation work from another branch, then re-reads claims whose code changed Free
bravos-report writes the assessment, one finding, a retest or an executive brief from gathered evidence Free
bravos-codegraph answers reachability, attack-surface and blast-radius questions Team
bravos-sast runs licensed static analysis, and says plainly when it is unavailable Team
bravos-ship the pre-PR security gate Team
bravos-cicd writes and audits the security GitHub Actions workflows Team
bravos-schedule /loop while you work, the nightly workflow, or a cloud task you create Team
bravos-campaign reaches a goal in rounds, such as coverage to 100% Team

Team features have a 30-day no-card trial. Without a licence, a paid skill stops at its first step and says so. On Free, bravos-threat-hunt reads the code and says plainly that the graph and SAST were unavailable.

Run bravos watch in a second terminal pane beside the agent. It is read-only and runs no model. It redraws as files change.

Terminal window
bravos watch --once
╭─ OPEN ──────╮ ╭─ HIGH+ ─────╮ ╭─ CONFIRMED ─╮ ╭─ TESTED ────╮ ╭─ COVERAGE ──╮ ╭─ STALE ─────╮
│ 2 │ │ 2 │ │ 0 │ │ 0/2 │ │ 100% │ │ 0 │
╰─────────────╯ ╰─────────────╯ ╰─────────────╯ ╰─────────────╯ ╰─────────────╯ ╰─────────────╯

Truncated to the metric panels, from the annotated audit-notes fixture. Further panels show recent changes, what to test next, the threat model, reports, setup, the code graph, the branch and team, top threats, SAST and the model’s size. --once prints one frame and exits; --json returns one bravos.watch/v1 snapshot.

Terminal window
bravos doctor
skills claude: 16 agents: 16
hooks claude: yes codex: yes cursor: yes
pre-pr claude: no codex: no (opt-in: bravos setup --pre-pr-hook)
autopilot claude: no codex: no (opt-in: bravos setup --autopilot)
cards claude: yes codex: yes
coach claude: yes codex: yes

Truncated to the repository’s skills and hooks. Above them doctor lists each tool with its path and the plan line; below them it checks for dead user-level hooks and old Python bravos skills, the last update check, the colour theme, and the two proof-tool capability checks. doctor is read-only. bravos doctor --fix-hooks removes Bravos’s own user-level hooks whose script is gone, after a backup.