Overplane: Safer AI Coding with Containers and Formal Verification

Skip to content Guide v0.0.8

Evolve Verified Software<br>Overplane<br>Safer AI Coding

Overplane combines popular AI coding agents with containers, spec-driven<br>development, and formal verification to create a safer and more reliable<br>AI coding workflow.

Free & open source · Apache-2.0 licensed<br>Get started Learn more first

Safer code ◆ Auditable codegen ◆ Model agnostic

The prop<br>Move fast without breaking your things.

Coding agents earned their reputation for speed, and for the occasional<br>catastrophe. Overplane keeps the speed and fixes the rest in three moves:<br>contain the agent, wrap it in a real build system, and verify the plan<br>before the code exists.

Your laptop is not a sandbox

One agent deleted a production database during a code freeze. Another destroyed a user&rsquo;s files while &ldquo;reorganizing&rdquo;<br>them. Even the agent vendors now say unattended runs belong in a container behind an egress firewall.

The fix

In Overplane the container is not a best practice you set up later.<br>It is the only way agents run: every agent, every build, the same<br>locked-down container with a restricted view of your system. A bad<br>day means deleting a sandbox, not restoring a backup.

A real build system for specs

Most spec-driven tools are Markdown prompts riding inside your<br>coding agent. Run one twice and you get two different codebases,<br>with no record of which run produced what. There is a growing case<br>that code generation now sits in the trusted computing base , where determinism and auditability stop being optional.

The fix

Overplane treats codegen like compilation: content-hashed sandbox<br>images, content-addressable outputs, cached incremental builds you<br>can replay down to a single spec, and a normalized dollar cost for<br>every agent run. A real build system, where the compiler happens to<br>be an agent.

Review the plan, not the 4,000-line diff

96% of developers don&rsquo;t fully trust AI-generated code, and over a third say reviewing<br>it takes more effort than reviewing a colleague&rsquo;s. Generated diffs<br>are the new toil.

The fix

Review moves upstream, where it is cheap. You sign off on short<br>numbered specs, Z3 checks them for contradictions before any code is<br>written, and the checked IR can keep going: in the rustdis example it becomes 48 passing property tests and Kani proof harnesses guarding<br>a wire-compatible Redis clone.

What is it?<br>Small binary, batteries included

Overplane is a small, single-file, open-source binary that orchestrates<br>recipes, prompts, containers, and AI coding agents into an opinionated<br>workflow:<br>plain-language specs in, working and verified software out.

Critical review: Dismissive Dan’s review of Overplane Spec-driven<br>Numbered Markdown specs are the source of truth. Schema-validated, git-diffable, and the only input the build consumes.

Sandboxed agents<br>Every coding agent runs in an identical local container with a restricted view of your system. OS-level isolation ensures reproducibility and security.

Replayable builds<br>Content-hashed images and content-addressable outputs make builds cacheable, auditable, and replayable down to an individual spec.

Spec validation<br>Raise lifts each spec into an Intermediate Representation; verify runs Z3 on per-spec and merged models before any code is written.

Richer codegen<br>Codegen gives the agent both the checked IR and the original spec, then reconciles the generated output only after a successful run.

Agent agnostic<br>Use your own API keys with Claude Code, Codex, or OpenCode, and choose the agent per project or per run.

Multi-agent support<br>Bring your favorite coding agent

Overplane is not another agent, and it is not another subscription. It<br>drives the headless agents you already pay for, on your own API keys, and<br>lets you swap them per project or per run.

Every agent below installs from official channels into a container image<br>you build, audit, and control.

OpenAI Codex<br>Anthropic Claude Code<br>OpenCode

Main features<br>How Overplane compares

The closest cousin is Spec Kit , and it is genuinely good: prompts that bring spec discipline to the<br>agent you already run. The difference is what happens after the spec. Spec<br>Kit hands your agent a better prompt; Overplane runs the whole build, and<br>owns the sandbox, the cache, the verification step, and the bill.

Capability Overplane Spec Kit 0.12 Kiro IDE 1.0 / CLI 2.12 Spec-driven authoring<br>Numbered, schema-validated specs are the durable source of truth for what you are building.

Yes Yes Yes Container-isolated agents<br>Every AI agent runs sandboxed in its own container, never touching your host.

Yes No No Vendor-neutral multi-agent<br>One abstraction over Claude Code, Codex, and OpenCode. Pick your agent per project or per run.

Yes Yes No Unified cost reporting<br>Normalized token usage and dollar cost for every agent run, aggregated automatically per build.

Yes No No Cached, hashed builds<br>Deterministic project images and...