arccode is a multi-provider agent harness that routes each task to the right model by complexity, cost, performance, and intent. Spawn specialist agents, load and build skills, and drive tools, Anthropic, OpenAI, and local Ollama behind one interface.
Everything you need to run agents across providers, decide which model fits, and extend the system at runtime.
Anthropic, OpenAI, Ollama (local), and OpenRouter behind one normalized interface. Models are hot-swappable.
A classifier scores every model by capability, cost, and latency for the task's intent and complexity. Overrides always win.
Markdown + frontmatter, auto-discovered. Each agent pins a model policy, tool allowlist, and skills.
Progressive disclosure: skill descriptions are always indexed, bodies load only when needed. Import or build new ones.
A coordinator spawns specialists and fans out subtasks, each routed to its own model.
Connect MCP servers, gate tools with hooks, and resume persistent sessions across runs.
arccode picks a model tier from the task's intent and complexity. Cheap models for bulk work, frontier models for hard reasoning.
| Intent | Model tier | Why |
|---|---|---|
| bulk read / summarize | small / cheap | high volume, low stakes |
| implement | mid workhorse | good tools + code, moderate cost |
| design / debug / review | frontier | needs strong reasoning |
| chat | small | latency matters |
arccode whichmodel "refactor the distributed cache layer" explains any decision.
Install, set a key (or run fully local with Ollama), and go.
# install (isolated, recommended) pipx install git+https://github.com/acnologiaslayer/arccode # or into the current environment pip install git+https://github.com/acnologiaslayer/arccode # pick your providers export ANTHROPIC_API_KEY=... export OPENAI_API_KEY=... # Ollama runs locally at :11434, no key needed # run a task (model auto-selected by the router) arccode run "Add a --json flag to the export command and test it" arccode run "Design a rate limiter for 10k rps" --agent architect -v arccode chat # interactive REPL arccode run "Start a review" -s new # resumable session arccode agents # list agents arccode models # catalog + pricing arccode whichmodel "fix the failing build"
Six specialists ship by default, each pinned to a sensible model tier. Add your own by dropping a Markdown file in the registry.
Understands the task, does it or decomposes and delegates. auto
Writes and edits code, verifies the change. workhorse
Designs systems, weighs trade-offs, diagrams. frontier
Root-cause analysis, reproduces and fixes. frontier
Bulk reading and summarization, cheaply. bulk-read
Correctness, security, and design review. frontier
A clean loop: classify, route, run tools, repeat, spawn when it helps.
Infer intent and complexity from the task.
Score every model by capability, cost, and latency.
ReAct loop: the model calls tools, arccode executes them.
Delegate subtasks to specialists, each self-routed.