Skip to content

Implementing oh-my-opencode-slim in Opencode

oh-my-opencode-slim turns Opencode into a multi-agent suite: one orchestrator plans the work, then auto-delegates to specialists. Mix any models from any providers, balance quality / speed / cost.

My setup runs entirely on free-tier models — no usage cost.

What you get

  • Orchestrator — master delegator, plans and reconciles work
  • Explorer — fast codebase reconnaissance
  • Oracle — architecture, hard debugging, code review
  • Librarian — external docs / web research
  • Designer — UI/UX implementation and polish
  • Fixer — fast bounded implementation
  • Observer — visual analysis (screenshots, PDFs, diagrams)

Opencode Multi Agent Suite · Mix any models · Auto delegate tasks

1. Install Opencode

Follow the official install: opencode.ai.

# quick check
opencode --version

2. Install oh-my-opencode-slim

bunx oh-my-opencode-slim@latest install

No Bun? npx works too since the CLI is a Node-compatible bundle:

npx oh-my-opencode-slim@latest install

This registers the plugin in ~/.config/opencode/opencode.json and creates ~/.config/opencode/oh-my-opencode-slim.json.

Docs: Installation Guide · Configuration.

3. Get API keys from providers

My preset below mixes free models from a few providers. Grab keys from each one you want to use:

  • Opencode Zen (opencode/*) — from your Opencode account
  • OpenRouter (openrouter/*) — openrouter.ai
  • Bai (bai/*) — Bai API console
  • Pollination (pollination/*) — pollination.ai

You don’t need all of them on day one — start with one or two, add the rest later. Each agent has a fallback list, so if the first model fails / rate-limits, the next one is tried.

4. Connect providers in Opencode

Inside Opencode:

/connect

Find the provider in the list and paste your key. If your provider isn’t listed, add it as a custom provider — see Opencode providers docs.

Repeat for every provider you collected keys for in step 3.

5. Refresh models

Once providers are connected:

opencode models --refresh

Then list to confirm Opencode sees them:

opencode models --list

If a model is missing here, it will be missing in the slim config too. Fix auth first, then continue.

6. Edit the slim config

Open the config file:

nano ~/.config/opencode/oh-my-opencode-slim.json

Here is my working opencode-zen-free preset — every agent on a free model:

{
  "$schema": "https://unpkg.com/oh-my-opencode-slim@latest/oh-my-opencode-slim.schema.json",
  "preset": "opencode-zen-free",
  "fallback": {
    "enabled": true
  },
  "disabled_agents": [],
  "presets": {
    "opencode-zen-free": {
      "orchestrator": {
        "model": [
          "opencode/muse-spark-1.3-contributor-free",
          "openrouter/thinkingmachines/inkling:free"
        ],
        "variant": "xhigh",
        "temperature": 0.4,
        "color": "#CBA6F7",
        "skills": ["*"],
        "mcps": ["*", "!context7"]
      },
      "oracle": {
        "model": [
          "bai/qwen3.8-flash",
          "opencode/muse-spark-1.3-contributor-free"
        ],
        "variant": "xhigh",
        "temperature": 0.3,
        "color": "#F38BA8",
        "skills": ["simplify"],
        "mcps": []
      },
      "explorer": {
        "model": [
          "bai/qwen3.8-flash",
          "openrouter/thinkingmachines/inkling-small:free"
        ],
        "temperature": 0.2,
        "color": "#89B4FA",
        "skills": [],
        "mcps": []
      },
      "librarian": {
        "model": [
          "openrouter/thinkingmachines/inkling-small:free",
          "openrouter/nvidia/nemotron-3-ultra-550b-a55b:free"
        ],
        "temperature": 0.2,
        "color": "#A6E3A1",
        "skills": [],
        "mcps": ["context7", "gh_grep"]
      },
      "designer": {
        "model": [
          "bai/glm-5.3-flash",
          "pollination/morriszdweck/glm-fast"
        ],
        "temperature": 0.3,
        "color": "#F5C2E7",
        "skills": [],
        "mcps": []
      },
      "fixer": {
        "model": [
          "opencode/muse-spark-1.3-contributor-free",
          "bai/glm-5.3-flash"
        ],
        "variant": "xhigh",
        "temperature": 0.3,
        "color": "#FAB387",
        "skills": [],
        "mcps": []
      },
      "observer": {
        "model": [
          "openrouter/thinkingmachines/inkling-small:free",
          "openrouter/nvidia/nemotron-3-ultra-550b-a55b:free"
        ],
        "variant": "low",
        "temperature": 0.2,
        "color": "#94E2D5",
        "skills": [],
        "mcps": []
      }
    }
  }
}

What matters in this file:

  • preset — which preset is active. Must match a key under presets.
  • fallback.enabled: true — if the first model in an agent’s list fails, try the next.
  • model as an array — priority order. First = primary, rest = fallbacks.
  • disabled_agents: [] — empty means Observer stays enabled (needed for images/PDFs since some orchestrator models aren’t multimodal).
  • skills / mcps per agent — permission grants. Orchestrator gets ["*"], specialists get only what they need. Librarian gets context7 + gh_grep because that’s its job.

Save, then restart Opencode.

7. Verify

opencode

Then in the TUI:

ping all agents

You should see every agent reply. If one fails, check its provider auth (step 4) and that opencode models --list contains its model (step 5).

8. Run it

Just describe the task — the orchestrator builds the work graph and dispatches specialists in the background:

Add dark mode to the settings page, check the Tailwind docs for the current way, and keep it responsive

That one sentence fans out to Librarian (docs) + Explorer (codebase) + Designer (UI) + Fixer (implementation), reconciled back into one result.

Manual control when you want it:

@explorer where is auth handled?
@librarian how does Opencode custom provider auth work?
@oracle review this refactor plan
@designer polish this landing page
@fixer rename getCwd to getCurrentWorkingDirectory across src/

Other useful bits:

  • /preset — switch the whole team’s models at runtime
  • run codemap — hierarchical map for unfamiliar repos
  • Bundled skills: deepwork, verification-planning, simplify, worktrees, clonedeps, reflect — see Skills docs

Troubleshooting

  • Agent model not found → you skipped step 5. opencode models --refresh, confirm with --list.
  • Auth error on one agent → /connect again for that provider, then refresh.
  • Want latest plugin code → run from master per the README, or pin a version in the plugin array.
  • Full reference → Configuration · Background orchestration.

That’s it — install Opencode, install the slim plugin, connect providers, refresh models, paste the config, and you have a free multi-agent coding team.