superclaw is a minimal Go agent runtime — bounded steps, curated tools, Docker-first isolation. It does the task and stops.
Minimal. Bounded. Container-first.
Designed with hard constraints so the agent cannot surprise you.
Step limits, wall-clock timeouts, and per-run URL caps. The agent cannot run forever. Every execution has a hard ceiling.
fetch_url, web_search, read_file, write_file, patch_file, list_files, run_bash. No more, no less. Each tool has a typed contract.
summarize, research, extract, coding, github. Injected as system prompt instructions, not as extra tools. Scope is explicit.
Non-root user, read-only rootfs, capability drops, /workspace bind mount. The container model is not optional — it is the design.
Exponential backoff with jitter on rate limits and server errors. Fetch retries too. Transient failures do not abort the run.
Every run appended to .superclaw/runs.jsonl. Inspectable, auditable history. No black box — every step is logged.
From zero to running agent in under two minutes.
Requires Go 1.23 or later. Installs the superclaw binary to your GOPATH.
go install github.com/akshaymemane/superclaw/cmd/superclaw@latestsuperclaw uses the Anthropic API. Get a key at console.anthropic.com.
export ANTHROPIC_API_KEY=sk-ant-...Pass the task as a quoted string. The agent plans, executes tools, and stops when done.
superclaw "read the README and write a summary to summary.md"Drop a superclaw.json in your project root to set defaults, restrict tools, or load a skill.
1{2 "model": "claude-opus-4-5",3 "max_steps": 20,4 "max_tokens": 4096,5 "timeout_seconds": 120,6 "work_dir": "/workspace",7 "max_fetch_calls": 5,8 "tools": [9 "read_file",10 "write_file",11 "list_files",12 "run_bash"13 ],14 "skills": ["summarize", "coding"],15 "system_prompt": ""16}All fields are optional. Defaults are conservative — 20 steps, 120s timeout, all tools enabled.
View all configuration optionsA simple, transparent execution model. No magic, no hidden state.
Reads superclaw.json (or defaults), validates the tool allowlist, and injects skill instructions into the system prompt.
Sends the task and system prompt to Claude. Receives a response with zero or more tool_use blocks.
Runs each approved tool call, checks against step and fetch budgets, enforces timeouts. Feeds results back to the model.
When the model returns a final text response (no more tool calls), writes the session to runs.jsonl and exits.
Max steps, max tokens, wall-clock timeout, and max fetch calls — all configurable, all enforced.