Skip to content

AI Agents

Use Bullpen CLI with Claude Code, Codex CLI, Gemini CLI, or OpenClaw to research markets, inspect your account, and run documented Bullpen workflows from natural language.

Setup

Install the CLI, check auth, preview one workflow, then install the agent skills:

bullpen status --output json
bullpen doctor auth --output json

Run bullpen login only when status or auth diagnostics say login is required. After auth is healthy, preview before confirming any live workflow:

bullpen polymarket search "bitcoin" --type market --limit 5
bullpen trade buy <MARKET_SLUG_FROM_SEARCH> No 5 --max-price 0.01 --preview --output json

Live: after reviewing the preview, live buy submission requires --yes:

bullpen trade buy <MARKET_SLUG_FROM_SEARCH> No 5 --max-price 0.01 --yes

Preview the install target first:

bullpen skill install --dry-run

Live: Install or update local AI-assistant skill files when you want local AI-assistant integration.

bullpen skill install

Restart your AI assistant after bullpen skill install. Agents load skills at startup, so an already-running Claude Code, Codex, Gemini, or OpenClaw session will not see newly installed or updated Bullpen instructions until it restarts.

Choose one assistant target. Preview that target, then install it.

Live: Each install command writes or updates local skill files for that assistant.

Claude Code:

Live: The second command writes or updates Claude Code skill files.

bullpen skill install --claude --dry-run
bullpen skill install --claude

Codex:

Live: The second command writes or updates Codex skill files.

bullpen skill install --codex --dry-run
bullpen skill install --codex

Gemini:

Live: The second command writes or updates Gemini skill files.

bullpen skill install --gemini --dry-run
bullpen skill install --gemini

OpenClaw:

Live: The second command writes or updates OpenClaw skill files.

bullpen skill install --openclaw --dry-run
bullpen skill install --openclaw

Use --dry-run or --check when you want to inspect or validate the target without changing files. Use --read-only when you want the installed skill to block mutating Bullpen commands.

What Agents Can Do

After the skill is installed and the assistant restarts, you can ask it to:

  • Discover and compare Prediction markets.
  • Check balances, positions, open orders, rewards, and points.
  • Preview or execute Polymarket-backed Prediction trades with bullpen trade.
  • Run bullpen fix --refresh only when status, diagnostics, or the failed command asks for refreshed readiness evidence. Run bullpen support bundle --output json when support needs a redacted artifact.
  • Recover stuck Polymarket funds or resolved positions with bullpen recover by following JSON fields such as recovery_plan.next_action, resolution_owner, safe_to_retry, and executable-leg evidence. Treat a next_command value as specific to the object that contains it.
  • Manage Perps (Hyperliquid) account, order, and position workflows exposed by the CLI.
  • Keep copy-trading venues separate: Polymarket copy trading lives under bullpen tracker copy; Hyperliquid copy trading uses the experimental bullpen hl copy namespace.
  • Use experimental Onchain (Solana) token reads, swaps, limit orders, DCA, and portfolio commands.
  • Fill March Madness brackets and use Bullpen bracket workflows.
  • Diagnose common setup and account issues without asking users to manually choose advanced wallet or exchange diagnostics first.

Money-moving commands still use the CLI's normal safety model: previews are available where supported, live execution requires the relevant command flags, and account state comes from your authenticated Bullpen session. If JSON reports auth_incident, outcome_unknown, mutation_outcome_unknown, requires_state_inspection, or safe_to_retry: false, inspect the emitted next command and current state before any manual retry.

For Prediction workflows, start with passive checks:

bullpen status
bullpen doctor auth

Then use the command that matches the task. For a trade, search and preview:

bullpen polymarket search "bitcoin" --type market --limit 5
bullpen trade buy <MARKET_SLUG_FROM_SEARCH> No 5 --max-price 0.01 --preview --output json

For funds recovery, preview the plan:

bullpen recover funds --dry-run --output json

When support asks for a redacted artifact:

bullpen support bundle --output json

Advanced Polymarket wallet and CLOB commands remain available in the reference, but they should be escalation commands. Use them only when the simplified command output or Bullpen support asks for that exact evidence.

Installed Skill Locations

bullpen skill install writes the bullpen-cli skill into detected assistant config directories. Personal installs use these paths:

Assistant Default skill path
Claude Code ~/.claude/skills/bullpen-cli/
Codex CLI ~/.codex/skills/bullpen-cli/
Gemini CLI ~/.gemini/skills/bullpen-cli/
OpenClaw ~/.openclaw/skills/bullpen-cli/
Cross-tool standard ~/.agents/skills/bullpen-cli/

When a non-Claude assistant is detected, Bullpen also writes the cross-tool standard copy under ~/.agents/skills/bullpen-cli/. Project installs use .claude/skills/bullpen-cli/ for Claude Code and .agents/skills/bullpen-cli/ for the other assistants.

Public Skill Manifest

The public mirror lives at docs/public/skill/SKILL.md, with supporting references under docs/public/skill/references/. It is generated from the maintained skill source and embedded into the CLI binary for bullpen skill install. Treat it as a published manifest, not a hand-edited user config file.

Seeing version: "0.0.0" in the public mirror is expected. Source-controlled skill manifests keep 0.0.0 as an install-time placeholder. During bullpen skill install and bullpen skill update, the CLI stamps installed copies with the running Bullpen CLI version so bullpen skill list can detect outdated local skills.

References