Cogo
An agentic CLI
for Go developers
Like Claude Code, but built in Go on the Google ADK and
Gemini 3.x. One static binary, MCP-native, project-scoped.
Two modes, one binary
Pipeable cogo -p for shell + CI; full Bubble Tea TUI on a TTY with streaming markdown, slash + @-file palettes, and prompt history.
Project-scoped
A .agents/ directory holds model, permissions, MCP servers, skills, and project memory. Auto-discovered like .git.
MCP-native
Drop a .agents/mcp.json describing stdio or HTTP MCP servers and their tools become callable. Schema-driven elicitation modal for servers that prompt the user.
Claude-compatible skills
SKILL.md bundles under .agents/skills/<name>/ are loaded lazily and invoked as agent tools. Same format as Claude Code.
Permissions you can trust
Three modes (ask / allow / yolo), in-TUI approval modal, non-overridable bash denylist, path-scope confinement for file tools.
Cost & telemetry
Per-turn ↑in · ↓out · $cost, session totals in the header, OpenTelemetry support, and JSON transcripts persisted on exit.
Try it in 60 seconds
# Auth (pick one)
export GOOGLE_API_KEY=... # public Gemini API
# or
gcloud auth application-default login # Vertex AI
export GOOGLE_GENAI_USE_VERTEXAI=true
export GOOGLE_CLOUD_PROJECT=your-project
# Single turn, pipeable
cogo -p "Summarize the files in this directory."
# Or interactive
cogoFor a fresh project: cogo init writes a sensible .agents/ skeleton; cogo init --interactive walks you through provider, model, and permission mode.