Native wire
OpenAI-compatible
Harness integration
DeepSeek Harness with Tokaroo
Run the DeepSeek harness against Tokaroo rather than tying it to one upstream model account.
Integration contract
Keep DeepSeek Harness's runtime. Replace the model connection.
Tokaroo does not host this harness. It receives the model request, selects a compatible route, handles fallback and metering, then returns the response in the harness's expected format.
Recommended model
Start with auto. Basic, Fast, Pro, and Max remain available for explicit workload requirements.
Family behavior
DeepSeek-first inside the selected quality band.
Profile status
The canonical Tokaroo wire and Gwen runtime path have passed the current compatibility contract.
Configuration
Install, configure, and run the native harness.
Use OPENAI_API_KEY for the Tokaroo credential and OPENAI_BASE_URL for the gateway. The recipe below is the same compatibility path exercised by the Gwen runtime.
Install or verify the harness
npm install -g @deepseek-ai/dsh@0.1.0-rc.7Shell environment
export TOKAROO_API_KEY="tok_..."
export OPENAI_API_KEY="$TOKAROO_API_KEY"
export OPENAI_BASE_URL="https://api.tokaroo.com/v1"
export TOKAROO_HARNESS_REF="deepseek-code"DeepSeek Harness provider patch - ./tokaroo-dsh.patch.json
[
{
"id": "agent-default-model",
"config": { "provider": "tokaroo", "model": "auto" }
},
{
"id": "llm-deepseek",
"disabled": true
},
{
"id": "llm-pi-ai",
"config": {
"providers": {
"tokaroo": {
"displayName": "Tokaroo",
"apiKeyEnv": "OPENAI_API_KEY",
"api": "openai-completions",
"baseURL": "https://api.tokaroo.com/v1",
"headers": { "x-tokaroo-harness-ref": "deepseek-code" },
"models": [{ "id": "auto", "maxTokens": 8192 }]
}
}
}
}
]Start with a connection check
export TOKAROO_API_KEY="tok_..."
export OPENAI_API_KEY="$TOKAROO_API_KEY"
export OPENAI_BASE_URL="https://api.tokaroo.com/v1"
dsh --profile headless --patch ./tokaroo-dsh.patch.json "Return exactly: Tokaroo connected"Required values
API key
OPENAI_API_KEY = your Tokaroo key
Base URL
OPENAI_BASE_URL = https://api.tokaroo.com/v1
Model
auto for the first run; switch to Basic, Fast, Pro, or Max only for a measured workload need
Profile
deepseek-code is a compatibility preference, never a hard model lock
The DeepSeek harness can use any Tokaroo-compatible model. Its native family is preferred only when it remains the best measured decision.
Verify the wire
Test Tokaroo before starting a full harness run.
This direct request proves the key, protocol, model alias, and harness profile header independently of the harness process.
OpenAI-compatible verification request
curl https://api.tokaroo.com/v1/chat/completions \
-H "Authorization: Bearer $TOKAROO_API_KEY" \
-H "x-tokaroo-harness-ref: deepseek-code" \
-H "Content-Type: application/json" \
-d '{"model":"auto","messages":[{"role":"user","content":"Return exactly: Tokaroo connected"}]}'Expected response
A successful native response containing Tokaroo connected.
Usage fields
Record the trace ID, requested Tokaroo mode, input and output tokens, latency, and Tokaroo charge.
Harness capabilities
CLI, headless, MCP, skills, plugins, subagents.
Production checklist
Keep credentials and execution boundaries explicit.
The harness remains responsible for its own installation, updates, sandbox, filesystem, tools, and network policy.
Store the Tokaroo key in a secret manager or trusted runtime environment.
Use Auto first, then choose another Tokaroo mode only for a measured workload requirement.
Preserve the harness's native wire and tool event format.
Log Tokaroo trace, requested-mode, token, latency, and charge headers with the harness run identifier.
Set spend limits, retry 429 and 503 responses with backoff, and monitor balance before production rollout.
Route DeepSeek Harness through one Tokaroo key.
Use Auto for the initial integration, then validate tool behavior, latency, usage, and output quality against your own workload.