A small, autonomous team-agent runtime — local-first, fast, written in Rust.
One main agent, up to nine workers, one durable journal. No RAG, no control plane, no permission engine, no evidence graph, no separate verification service.
minimal-agent gives a language model a real, autonomous team instead of a single conversational chat loop.
- Main Agent (Depth 0): Decomposes the mission, assigns bounded tasks, recalls workers, and synthesizes progress.
- Worker Agents (Depth 1): Up to 9 parallel workers executing discrete tasks and messaging Main and siblings directly.
- Durable Run Journal: A segmented, append-only ledger of raw events for zero-data-loss crash recovery.
- Self-Owned Briefs: Every agent curates its own
brief.md(current battlefield understanding) via semantic compaction.
- Simple is Best (curl philosophy): The smallest thing that works. A lean, sharp kernel that scales through composable CLI primitives, not monolithic bloat.
-
Prompt/Skill-First (Zero Code Bloat): Never solve in code what can be solved in prompts or skills. Code is a permanent maintenance liability. The Rust runtime strictly enforces OS/IO invariants (
bash, resource caps, process reaping), while domain tradecraft lives in modular skill prompts. -
Bounded Team Hierarchy: Fixed depth of 2 (Main
$\to$ Workers) prevents runaway recursive spawning and keeps token budgets strictly bounded. - Evidence Over Claims: Nothing is "done" simply because the model claims so. Completions, vulnerabilities, and flags require concrete tool output or executable verification.
Create a .env file in the project root:
OPENAI_API_KEY="your-api-key"
OPENAI_MODEL="your-model-name" # e.g., glm-5.3-flash, gpt-4o, claude-3-5-sonnet
OPENAI_BASE_URL="https://api.openai.com/v1" # optional (supports any OpenAI-compatible endpoint)
# OPENROUTER_API_KEY="your-key" # optional# Option A: Docker Compose (Recommended)
docker compose -f docker/compose.yaml run --rm minimal-agent
# Option B: Docker One-Shot CLI
docker run --rm -it --init \
--cap-add=NET_RAW --cap-add=NET_ADMIN \
--env-file .env \
-v ${PWD}/workspace:/workspace \
-v ${PWD}/runs:/state \
agnusdei1207/minimal-agent:0.110.0Current release: 0.110.0.
# Option A: Build and launch isolated Docker TUI
npm run check
# Option B: Global CLI installation
npm install --global minimal-agent@0.110.0
minimal-agent run --goal "Investigate the target and solve the objective" --workspace .MIT