For developers supervising Claude Code, Codex, and local coding agents

Keep coding agents inside
clear repo boundaries.

Voss gives local agent runs scoped tools, reusable project memory, approval gates, independent review, and an audit trail you can replay before you trust the result.

$npm i -g @vosslang/cli

See the audit trail

Bundles the Voss Python harness with vendored Python 3.12.

Voss vdiff output showing a .voss source file beside generated Python.

Six primitives, one org.

Voss models software work like a high-performing engineering organization — not a rigid automation pipeline. Every run is built from the same six primitives.

Capabilities

Toolbelt

Every tool is a typed, permissioned, auditable capability. Network and shell are default-deny unless a role is granted them.

Principles

Culture

Engineering principles are first-class config in .voss, injected into every agent context and recorded in the audit.

Orchestration

Delegation

An Engineering Manager loop turns one idea into scoped cards, assigns roles, partitions budget, and integrates the result.

Roles

Specialists

architect, backend, frontend, tester, reviewer, skeptic, docs — declared in .voss team{} with their own scope, budget, and model tier.

Memory

Knowledge

VOSS.md, project memory, session trees, and decisions keep institutional context inspectable across runs.

Verification

Review loop

Independent Reviewer-A and Reviewer-B gate completion. Agents cannot mark their own work done.

Engineering Manager loop

One idea in. Audited work out.

The Engineering Manager loop is the orchestrator: a constrained tech lead that decomposes, delegates, verifies, and integrates — and asks for you only when it matters.

Scope into cards

The EM converts one human idea into bounded work cards with acceptance criteria.

Assign roles

Each card is routed to a declared role from the team roster, with a recorded rationale.

Partition budget

Budget and scope fan out down the session tree. No child can overspend its parent.

Execute in parallel

Workers run concurrently inside their scope, within WIP limits, where it is safe to.

Verify continuously

Reviewer-A authors the verification bar from the original idea — not the EM's summary.

Review independently

Reviewer-B judges the diff narrative-blind and can fail idea-divergent work.

Block or integrate

Unverified or out-of-scope work is blocked with a reason. Clean work is integrated.

Audit and sign off

A replayable audit report is produced. Humans sign off only at meaningful moments.

The MVP flow

Declare a team. Run a goal.

A team is declared once in .voss: roles, scope, budget, and tools. Then a single goal runs as scoped, budgeted, independently reviewed work — with a replayable audit at the end.

How it works See the audit

voss team run

$ voss team check
  team "default" ok — 7 roles, budget 120k, scope src/** tests/** docs/**

$ voss team run "Add password reset flow with tests"
  → 4 cards derived · assigned architect, backend, tester, docs
  → budget partitioned · 4 workers dispatched
  → reviewer-A authored 6 checks · reviewer-B verdict: pass (0.91)
  → run_id 7f3a9c · 1 card blocked (rescoped) · audit ready

$ voss review 7f3a9c
  per-card reviewer-A + reviewer-B verdicts · evidence refs · outcomes

Three workflows. One control layer.

Run the canonical samples/ programs with voss run, or compare them with the raw Python equivalents in examples/raw_python/.

Confidence-gated intent classification.

$voss run samples/classify.voss

fn classifyIntent(input: string) -> string {
    let intent: probable<string> = ask("Classify the intent: " + input)

    if intent @ p >= 0.80 {
        return intent.value
    } else {
        return "unknown"
    }
}

The CLI.

One binary. Compiler verbs and agent verbs share a namespace, so you stay in flow whether you're shipping a program or asking it questions.

$ voss team checkValidate roles, scope, budget, tools
$ voss team run "add password reset"Run the goal as an engineering team
$ voss boardWatch cards move across the board
$ voss review <run_id>Inspect independent reviewer verdicts
$ voss session tree <root_id>Per-agent budget, scope, status
$ voss capabilities listInspect the agent toolbelt
$ voss recall "permission gate"Search code and project memory
$ voss do "summarize this PR"Single bounded agent task
$ voss chatInteractive REPL with live subagents
$ voss doctorDiagnose your environment

Install.

The npm package vendors Python 3.12 and the Voss wheel. Use pip when you manage Python 3.11+ yourself.