Super Intelligence Dog is the engine: one OpenAI-compatible API, open-weight models on measured routes with automatic failover, and bring-your-own-key for the closed-lab models you already pay for. Its agent runs the build loop — you approve the plan, it edits in a sandbox, an independent model re-reviews the change, and every result comes back as a receipt: what it cost and what it saved. BellerBuild is the proof layer that scans the result and issues the badge.
Reads and edits your real working tree, runs tools, keeps durable sessions, and never commits or
pushes on its own. Installs from PyPI with pipx or pip.
pipx install sid-harness # wheel from PyPI # first run — verify your key (/v1/me), save it, pick a default dir: sid setup sid run --task "Fix the off-by-one in add()."
Upload a project folder, approve the acceptance criteria and test command, and run in an isolated sandbox. You get the diff, test output, an independent review, a scan, and a receipt. Nothing publishes automatically.
Codex CLI, Cursor, Cline, OpenCode, Zed and anything OpenAI-compatible can use SID as their model backend — routed, metered, and capped under your key.
DeepSeek, Qwen, GLM, Kimi and more on measured routes with automatic failover. OpenAI and Anthropic models are bring-your-own-key — never resold, never a dependency. One OpenAI-compatible API for your own tools, plus a workspace for agent work (experts, projects, memory).
The agent proposes a bounded plan; you approve exactly what it may do. It edits the smallest change, runs your tests in a sandbox (never on a host), and an independent model re-reviews the diff. The receipt shows cost plus what the same work would have cost at closed-lab list rates — the saving is printed, not promised.
The BellerBuild review scans what actually leaks: database rules that expose rows, secrets in the tree, migrations that widen access. Every finding names the file and the fix. A verified build earns the badge anyone can check — and a paid plan can list it on the Beller Store.
Six steps — and a dog bred to smell out what everyone else walks straight past, at every one of them. You approve the plan; SID runs the rest and hands back a receipt. Nothing publishes on its own.
Every request is read for what it actually is — intent, domain, difficulty — then matched to the model that scores best on that task, not the one with the cheapest price tag. The pick is explained, and you can always override it.
Grounded: POST /v1/route returns the ranked candidates, a cost
estimate, and the reason it chose — the same pick the run uses.
The agent proposes a bounded plan and the acceptance criteria it will be judged against — then it stops. Nothing runs until you approve exactly what it may touch, and the task and files are locked after that line.
Grounded: the Build Workbench locks criteria and the test command at its human-approval step, before any run.
The agent makes the smallest edit that satisfies the approved plan, then runs your tests inside an isolated sandbox — never on the host. You get the diff and the test output, not a black box.
Grounded: the loop runs against SID's own endpoint inside a microVM sandbox and returns a unified diff plus the test outcome.
Before the change reaches you, a different model re-reviews the diff against the approved criteria. A weak patch gets caught here, not after it reaches your users.
Grounded: SID's verify mode runs a second model over the patch and returns a verdict alongside it.
The result is security-scanned for the things that get people breached: database rules that expose rows, secrets left in the tree, migrations that quietly widen access. Every finding names the file and the fix.
Grounded: the run calls the BellerBuild scan and fails closed if the scanner is unreachable; the receipt carries the findings.
Every run comes back as a receipt: what it cost, what the same work would have cost at closed-lab list rates, which model ran, and whether the checks passed. Nothing publishes on its own — the store is a next step you take, not one we take for you.
Grounded: the receipt carries cost, the saving, the diff, test output, review verdict, and scan findings; publishing is a separate, explicit action.