Install

1. Install

Python 3.10 or newer and uv. It gives you two commands, decosa and decosa-mcp.

Then save your dk_ key once. It asks for the key without echoing it, stores it with mode 600 and makes one test call, which returns a receipt.

2. Run the Decosa agent

• On your own GPU
  decosa agent --local http://127.0.0.1:8000/v1Copy
  Your vLLM (set up with the code kit at decosa.ai/apps/code), started with --enable-auto-tool-choice --tool-call-parser qwen3_coder. Nothing leaves your machine.
• Hosted on Decosa (receipted)
  decosa agentCopy
  Interactive; add -p "<task>" for one task. Don't send proprietary code to the hosted route. Needs Node 22.19+ (it fetches pi, pinned, the first time).
• Spend and receipts
  decosa usage && decosa verify <receipt-id>Copy

4. Ask it

• “Check this support answer against these two docs with Decosa, then verify one of the receipts.”
• “Can an open model take over my ticket-classification prompt? Which Decosa tool, and what does a run cost?”
• “Can I self-host Decosa's coding model on an RTX 4090?” (No: the pinned model needs a Blackwell card or 40 GB.)

The tools it adds

decosa_list_use_cases

Find the right tool for a job, with its measured cost and time per run

decosa_use_case

Its routes, the working example call, samples, limits and request shape

decosa_run_use_case

Run it on a sample or your input; saves the full output and returns the receipt ids

decosa_get_receipt / decosa_verify_receipt

Fetch a receipt and verify it offline against the pinned signing key

decosa_verify_record

Check a signed record (the hash-chained report most tools return)

decosa_verify_confidential

Check a confidential run's hardware evidence offline

decosa_selfhost_plan

Pinned images and models, whether your GPU fits (by model and architecture), and a smoke test

decosa_docs

Search the API contract: keys, errors, limits, receipts

How well it works

93%

of 46 Decosa tasks done right by Qwen3.8 with these tools (100% with thinking on)

46%

the same model as a generic agent with our docs and a raw HTTP tool

$0.007

per task at list price, a quarter of the model calls a generic agent makes

Our own eval, 46 tasks written for it, 28 Sep 2026: not a public benchmark. In Claude Code with the MCP server, 4 of 5 hand-run tasks were right the first time. The fifth (can an RTX 4090 run the coding model?) was wrong, and the tool now checks the GPU architecture. With no tools at all, the same model gets 22%.

Cost: the tools are free. Tool runs are billed per run, as on each tool's page. When the agent's own model is Decosa's hosted Qwen, it used about $0.007 per Decosa task at list price in the eval.