Configure OpenClaw with staik to use all models directly via Telegram, Discord, or the terminal.

Configuration

Add the following to your openclaw.json:

json

{
  "models": {
    "providers": {
      "staik": {
        "baseUrl": "https://api.staik.se/v1",
        "apiKey": "sk-st-your-key",
        "api": "openai-completions",
        "models": [
          {
            "id": "gemma4:31b",
            "name": "Gemma 4 31B",
            "contextWindow": 262144,
            "contextTokens": 64000,
            "maxTokens": 8192
          },
          {
            "id": "qwen3.6:35b-a3b",
            "name": "Qwen 3.6 35B A3B",
            "contextWindow": 262144,
            "contextTokens": 64000,
            "maxTokens": 8192
          },
          {
            "id": "qwen3.5:9b",
            "name": "Qwen 3.5 9B",
            "contextWindow": 262144,
            "contextTokens": 28000,
            "maxTokens": 4096
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "staik/gemma4:31b",
        "fallbacks": ["staik/qwen3.6:35b-a3b", "staik/qwen3.5:9b"]
      },
      "compaction": {
        "mode": "safeguard",
        "keepRecentTokens": 20000,
        "reserveTokens": 24000
      }
    }
  }
}

Switch models with /models in OpenClaw. fallbacks gives automatic failover if the primary model is unavailable. Which model fits which role — see Models.

Fine-tune context handling

Parameter	Where	Description
`contextTokens`	per model	Max tokens of conversation history per request. Lower if the model times out.
`maxTokens`	per model	Max tokens in the model response. Lower if responses get cut off.
`keepRecentTokens`	compaction	Tokens of recent messages that are never compressed.
`reserveTokens`	compaction	Tokens reserved for the response during compaction. Raise if responses get truncated.
`contextWindow`	per model	The model's total context window (262K for all chat models).

Quick diagnosis

Model stops responding mid-conversation → lower contextTokens. Responses get truncated → raise reserveTokens. Model forgets what you just said → raise keepRecentTokens.

Configuration per plan

Every request counts both prompt tokens (history) and completion tokens (response). The key to staying within budget is contextTokens — it controls how much history is sent per request.

json

{
  "models": {
    "providers": {
      "staik": {
        "baseUrl": "https://api.staik.se/v1",
        "apiKey": "sk-st-your-key",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen3.5:9b",
            "name": "Qwen 3.5 9B",
            "contextWindow": 262144,
            "contextTokens": 12000,
            "maxTokens": 2048
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": { "primary": "staik/qwen3.5:9b" },
      "compaction": {
        "mode": "safeguard",
        "keepRecentTokens": 6000,
        "reserveTokens": 8000,
        "maxHistoryShare": 0.3
      }
    }
  }
}

Rules of thumb per profile:

Profile	contextTokens	maxTokens	Model	Capacity (approx.)
Frugal (PAYG/EA pot)	12,000	2,048	qwen3.5:9b	~80 requests per 1M tokens
Hobby (250K/day)	24,000	4,096	9b + 35b	~15 rounds/day
Agent (500K/h rolling)	48,000	8,192	gemma4 + 35b + 9b	~10 requests/h

The fast 9b model uses the fewest tokens; a low maxHistoryShare compresses early and saves budget. With rolling hourly windows (the Agent plans), tokens from an hour ago no longer count.

Sub-agents (ACP)

For larger projects (e.g. a multi-file site), OpenClaw splits the work into sub-agents via ACP (Agent Coordination Protocol). Each sub-agent gets isolated context — without ACP the main agent must hold the entire project state in every request (5 files × 30K context = 150K+ tokens in a single round); with ACP token usage can drop 70–80%.

json

{
  "agents": {
    "list": [{
      "id": "<your_agent_id>",
      "runtime": {
        "type": "acp",
        "acp": {
          "agent": "codex",
          "backend": "acpx",
          "mode": "persistent"
        }
      }
    }]
  },
  "session": {
    "threadBindings": {
      "enabled": true,
      "spawnSubagentSessions": true
    }
  }
}

Verify in your chat: /acp spawn <agent_id> spawns a sub-agent manually, /verbose shows in real time which sub-agents are spawned and where errors occur.

Gateway connection closed?

The error "Cannot start sub-agent (gateway connection closed)" refers to OpenClaw's internal gateway, not the staik API. Check that the acpx backend is installed, gateway.mode: "local" is set, the gateway runs on the right port, and no stale processes block it. When ACP fails the main agent does everything itself — the entire project context is sent in every request and tokens burn fast.

Troubleshooting

OpenClaw produces no response

Turn off streaming in the provider block — some channels have issues with streaming responses; without streaming a regular request is sent that waits for the full response:

json

{
  "models": {
    "providers": {
      "staik": {
        "baseUrl": "https://api.staik.se/v1",
        "apiKey": "sk-st-your-key",
        "api": "openai-completions",
        "streaming": "off",
        "models": ["..."]
      }
    }
  }
}

Agent asks "Should I run?" between steps

The behavior comes from OpenClaw's own agent prompt, not from staik. Override it with a system prompt — per Telegram group via channels.telegram.groups.<chat_id>.systemPrompt, or globally via the workspace files (~/.openclaw/workspace/IDENTITY.md or SOUL.md):

text

You are an autonomous agent. Execute ALL steps of a task in sequence without
asking between them. Stop only on (1) an explicit error requiring a decision,
or (2) when the entire task is done.

Other causes: the sub-agent could not start (see ACP above), or the task requires more output than maxTokens — raise it to 8192 if budget allows.

429 mid-session

You've hit your token limit. Lower contextTokens (64000 → 16000–32000), switch to qwen3.5:9b as primary, or upgrade your plan — see Error codes and rate limits. Monitor in real time via the X-RateLimit-Used-Tokens header.