open models

Run more agents without paying frontier rates for every token. Set up your open models once, then let crystl route each job by capability, cost, privacy, and available hardware.

works withany CLI agent

Keep the expensive thinking where it belongs. Use a frontier model for architecture, difficult debugging, and the decisions that need your strongest agent. Put lint fixes, test runs, migrations, documentation passes, and other token-heavy work on an open model running on your hardware or through a managed host such as z.ai.

Set it up once. Connect Ollama, LM Studio, vLLM, llama.cpp, or a compatible hosted service in Settings → agents → providers & keys. crystl checks whether the model is reachable, keeps any per-connection key in your Keychain, and supplies the right flags and environment variables when a new shard starts.

Orchestrate the whole mix. A lead agent can keep the hard problem on a frontier model while sending bounded work to open-model workers. Every worker remains a normal crystl shard, so you can watch it, steer it, and bring its result back into the larger job.

Open models, also called open-weight models, give you control over where and how they run. They are not automatically private or equally capable. A model and agent both running on your hardware can keep the work local. A managed host still receives your code, and a local worker managed by a hosted lead is not an air gap. Match the task to the model and the savings become useful instead of merely cheap.

Why it matters

Set up once

Connect a local runtime or hosted open-model service once. crystl handles discovery, credentials, agent-specific flags, and each new shard's environment.

Orchestrate every model

Mix frontier and open-model workers in one job. Route each task by difficulty, cost, privacy, and the hardware you have available.

Cut token costs

Move repetitive, high-volume work off metered frontier APIs. Pay premium token rates only for the judgment that needs them.

Know what is ready

See whether each connection is reachable and which Ollama models are warm. A missing or unavailable model fails before a worker opens.

Three steps from model to worker

Start the model, give the connection a name, then choose it when you open a shard. crystl handles the agent-specific plumbing. The optional key stays in your Keychain, and anything you export yourself still wins.

Settings → agents → providers & keys → open models

  label                  closet-box
  kind                   Ollama
  base URL               http://100.x.y.z:11434     ← as seen from your Mac
  as seen from remote    http://localhost:11434     ← for agents running on the box
  key                    (none)
  env vars               OLLAMA_API_BASE, OLLAMA_HOST
  enabled                yes

  label                  z-ai
  kind                   anthropic-compatible
  base URL               https://api.z.ai/api/anthropic
  key                    ••••••••••••            ← Keychain, never your Anthropic key
  env vars               ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN
  enabled                yes
    

New shards pick this up at launch. A shard that was already running keeps the old values until you restart it. See the open models setup guide for the full walkthrough.

Orchestrate by model size

Different agents spell model ids differently. Instead of looking each one up, let crystl read what the connection serves and propose a small, standard, and large model. Review the mapping once, then heroes, quests, and orchestrators can choose by size.

$ crystl agent profile suggest --agent codex --slot closet-box
closet-box serves 6 models. proposed profile for codex:
  small      gpt-oss:20b
  standard   qwen3-coder:30b
  large      qwen3-coder:480b
write this profile? [y/N]

aider      ollama_chat/qwen3-coder     # ollama_chat beats ollama here
opencode   ollama/qwen3-coder
    

That one map turns model choice into an orchestration control. Run lower-cost workers at small or standard, reserve large for harder work, and change the actual model later without rewriting every hero or workflow.

Where open models pay off

High-volume mechanical work

Send lint fixes, test loops, migrations, documentation passes, and other bounded tasks to an open model. Keep the frontier model focused on decisions that justify its token price.

A mixed-model agent team

Let a strong lead agent plan the work, then hand well-scoped pieces to self-hosted or managed open models. crystl keeps every worker visible and gives each result a clear path back.

Your own hardware

Run the agent and model on a Mac, workstation, or GPU box you control. A remote-vantage URL keeps the same model available from your desk and an SSH session on the machine.

Hosted open models

No GPU required. A managed host such as z.ai runs the model for you, through the same shards, model sizes, and orchestration controls.

Setup tutorials

Each one goes from nothing installed to a working shard on that server.

Dive deeper in the Open models setup docs

Ready to try it?

Create a free account and download crystl for macOS.

get started getting started guide →

Frequently asked questions

Book of Spells
the crystl guild

Crystl's Book of Spells

Master crystl with a grimoire of powerful prompts. Summon heroes, assemble a party, revive shards, and complete epic quests.

open the book →

More features

Gems & Shards

Gems, shards, and the crystal rail keep every project one click away.

Action Panels

See what every agent is doing at a glance.

Agent Activity Panel

One live view of every agent across all your gems: who's working, who's idle, and where tokens are going.

Schedule Agents

Launch a fresh agent at a set time (once, daily, or weekdays). Pick the gem, agent, model, and prompt; crystl fires it on schedule.

Parallel Sessions

Run multiple Claude instances on the same repo in separate git worktrees.

Hero Shards

Summon a single hero as a shard (its own agent, model, and persona) without a quest party.

History Navigator

Browse, jump through, and search every session: conversation and tool calls.

Token Tracking

Track token burn per shard, estimate turns left, and reclaim context window by disabling heavy plugins and skills.

Workbench

A slide-out task list in every project, backed by a WORKBENCH.md the agent reads and writes with you.

Markdown Editor

Render and edit any .md in a side panel. Toggle it from the workbench, and let agents open docs for you from the CLI.

Prompt Integration

Facet Inserts give you one-click access to saved prompts, commands, and shortcuts.

Agent File Editor

Create, edit, and manage CLAUDE.md, rules, skills, and agent config files with a block editor, file chooser, and project optimizer.

Project Optimizer

Scan your project for gaps in agent configuration (missing files, incomplete docs, oversized code, and setup issues) with one-click fixes.

Skills Library

Browse and import agentic skills that teach agents like Claude Code or Codex new tricks: code review, testing, deployment, and more.

Cavrn

The terminal as structured data: every turn, tool call, result, and table captured as agent-readable metadata. Drawn fast on the GPU with Metal.

Remote Development

Full crystl integration over SSH: approval panels, file relay, and click-to-open on remote machines.

Agent Orchestration

One agent spawns, supervises, and merges the work of many: a control room for a team of coding agents.

crystl quest

Assemble a party of specialized agents. Each brings unique skills, coordinates in shared chat, and executes in parallel.

Screenshot Bar

Drag recent screenshots into the terminal. Spotlight-powered, always up to date.

Copy Bar

A one-click copy bar under the terminal. Agents stage text, you click Copy. Editable, up to 10 tabs, free.