Skip to content

Build your first threat model

Set Bravos up in a repository, open your agent there, and have it map the first threats. Everything on this page works on the Free plan.

  • Bravos installed, with GuardLink on PATH (bravos setup --check shows ✔ guardlink).
  • A git repository with at least one commit.

In a terminal, from the repository root, run bravos setup and answer the repository’s questions. To name the agents instead of being asked:

Terminal window
bravos setup --agent claude,codex
Harnesses: claude, agents
guardlink: initialised (guardlink init --agent claude,codex)
Skills: 56 written, 0 updated, 0 unchanged, 0 skipped
Guardlink skills mirrored to .agents/skills: 11 written, 0 updated, 0 unchanged, 0 skipped
Hooks: .claude/settings.json registered
Hooks: .codex/hooks.json registered
MCP: .mcp.json added cxg; kept guardlink
MCP: .codex/config.toml added guardlink, cxg; kept none
MCP deferred: codegraph — bravos setup --codegraph fetches codegraph-build and codegraph-mcp (Team plan); then bravos graph build .
In Codex run /hooks once to trust .codex/hooks.json
Next: open a session and say anything. Bravos begins by learning what the app is for, then maps the threats in your code.

Truncated to Bravos’s summary. GuardLink’s own initialization output comes first. The capture is from audit-notes, a one-file Flask fixture with a note search route and an export route.

When .guardlink/ is absent, setup runs guardlink init first. It then writes the 16 Bravos skills for each agent, the session hooks and the MCP wiring. New repositories use GuardLink’s default annotation mode, external .gal sidecars, unless you pass --guardlink-mode inline. A repository with a .cursor/ folder also gets Cursor’s rule and session hook; --agent cursor adds them without one.

Commit what setup wrote, so every teammate who clones the repository gets the same setup. Share Bravos with your team lists the files and what a teammate sees.

Open your coding agent in the repository and say anything. In Codex, run /hooks once to trust .codex/hooks.json. The session-start hook gives the agent the repository’s security picture, so no command is needed.

With no recorded intent, the agent first drafts what the app is for from the README, docs and code, then annotates against that draft. It opens a page for you to confirm the draft, and does not wait for you to do so. Give it your design documents or an earlier threat model to start from. See context from your design documents.

The agent writes GuardLink annotations and runs guardlink lint until it reports no errors. It asks before the first commit of a session. An @exposes it writes is still a hypothesis: only an authorized run against a target you own confirms it. See Annotate a repository and the verification loop.

Terminal window
bravos next
Setup 1/9 for audit-notes
Plan: Free (researcher) — the pre-PR gate, CI and nightly workflows, scheduled upkeep, multi-repository workspaces, campaigns, SAST, the code graph are on Team · bravos account activate starts a 30-day Team trial, no card
✘ build the code graph
unlocks: every later step reads it: real routes for intent, reach for annotate, call paths for templates, callers for rechecks
how: the code graph is on the Team plan: bravos account activate starts a 30-day trial (no card), then say "build the code graph"
→ record what the app intends
unlocks: severities judged against what the software is meant to do and which attackers are in scope
how: ask "learn what this app intends" (/bravos-context)
✔ cover the code (now 100%)

Truncated after the coverage step. The fixture’s two routes are annotated, so coverage is 100% and the score is 1 of 9. The arrow marks the next step. On Free, the code graph and SAST stay listed as Team steps, and the research loop goes on without them. bravos next --json prints the same as bravos.setup/v1.

Terminal window
bravos setup

In a repository that already has Bravos, setup brings its skills, hooks and wiring up to the installed version. In a session, the agent does this for you when the briefing says the repository is behind. See Upgrade Bravos.

Use Bravos from an agent covers what each agent gets and how to run a skill.