Skip to main content

​
Supported Agents

Gump orchestrates coding agent CLIs. Each agent runs in headless mode inside the worktree. Gump prepares the context, launches the process, streams its output, and collects the result.

​
Agent table

AgentCLI commandModelsContext fileResumeTokensNative cost
Claude Codeclaude -pclaude-opus, claude-sonnet, claude-haikuCLAUDE.md--resume <uuid>Per messageYes
Codexcodex execcodexAGENTS.mdresume <thread-id>Per turnNo
Gemini CLIgemini -pgemini, gemini-pro, gemini-flash, gemini-flash-liteGEMINI.md--resumeAggregatedNo
Qwen CLIqwen -pqwen, qwen-plus, qwen-maxQWEN.md--resume <uuid>Per messageNo
OpenCodeopencode runopencodeAGENTS.md--session <ses_id>Per step_finishNo
Cursorcursor -pcursorCLAUDE.md + .cursorrules--resume <uuid>Per messageYes

​
Using agents in workflows

Reference agents by name in your workflow YAML: 
- name: impl
  agent: claude-sonnet
For agents with multiple models, the name maps to a specific model flag. For example, claude-sonnet tells the Claude Code CLI to use Sonnet. If you just write claude, the CLI uses its default model. Check which agents are available on your machine: 
gump models
This lists all supported agents and whether their CLI is installed.

​
Connector roadmap

WaveAgentsStatus
Wave 1Claude Code, Codex, Gemini, Qwen, OpenCodeAvailable
Wave 1.5Cursor CLIAvailable
Wave 2Aider, Copilot CLI, Amp, Kimi CLIPlanned
Wave 3Trae CLI, Neovate, FetchcoderEvaluating
Any agent CLI with a headless mode and file-based output is a candidate for a Gump adapter.

​
Compat mode

If an agent provider changes its output format or breaks its CLI, Gump falls back to compat mode: raw stdout (not parsed for metrics), git diff as the source of truth for file changes, and fresh sessions only. The workflow still works — you just lose granular metrics until the adapter is updated.