BoundedCode

A local-first control plane for AI-assisted software engineering on large, long-running, multi-repository projects. It runs on commodity hardware and uses frontier models only for high-value escalation.

Local models do most of the work. The durable value is the system around the model: task continuity, repository intelligence, deterministic verification, sandboxing, routing and cost control. Models are replaceable configuration.

Status: pre-alpha. See docs/development/status.md for what is implemented, experimental and planned. Benchmark numbers in this repository are measured on the reference machine and linked to their raw reports. Nothing is claimed without a measurement.

How it fits together

boundedcode CLI (Go control plane)
 ├─ task ledger + audit log (SQLite)
 ├─ git worktrees  agent/<task-id>
 ├─ verification engine (gofmt, go test, tsc, eslint, gitleaks, ...)
 ├─ repository intelligence ──> codebase-memory-mcp (external): graph, impact (breadth)
 │                               + cross-service contract analyzers (HTTP, Kafka, env, Terraform)
 │                               + Serena v1.7.0 (optional, MIT-pinned): LSP symbols per worktree (depth)
 ├─ agent runtime ── JSON-RPC/stdio ──> OpenHands SDK adapter (in a container)
 │                                         └─ LLM calls tunnelled back to Go
 ├─ inference runtime ──> llama.cpp llama-server (external, supervised)
 └─ frontier gate (Z1-Z4) ──> codex exec with ChatGPT sign-in (optional)

Read docs/architecture/system-architecture.md for details and docs/product-spec.md for goals and constraints.

Measured so far

These numbers come from a single laptop (see below), with every result linked to its raw report. They come from 11 synthetic tasks, so they are indicative, not general claims.

	Qwen3.6-35B-A3B	Laguna XS 2.1	Qwen3-Coder-Next (3-bit)
Decode t/s at 2K / 58K context	39 / 30	~38 / ~31	27 / 21
Engineering suite, local only (hidden checks)	10/11	11/11	10/11
Hard cross-service task (4 runs)	1/4	4/4	0/1
Verified tasks per hour	18.2	7.0	9.8

• Milestone 1: a multi-repo task survived a forced process kill and two context condensations and completed locally (report).
• Frontier escalation (Codex, ChatGPT sign-in) fixed the hard task for Qwen3.6 with 2 messages (report).
• Cross-service contract analysis: an event-field rename that names only the producer passed hidden checks 3/3 with the analyzers and 0/3 without them. All three runs without them passed per-repo tests while breaking the consumer (evaluation).
• Details: infrastructure reports and model evaluation.

Quick start (development)

The full walkthrough (prerequisites with versions, model download, sandbox image, first task) is in docs/usage/getting-started.md.

make build
./scripts/build-llama-cpp.sh                 # pinned llama.cpp with CUDA (optional if you have one)
./bin/boundedcode init \
    --models-dir ~/models \
    --llama-server ~/.local/share/boundedcode/runtimes/llama.cpp/v0.5.0/bin/llama-server \
    --llama-bench  ~/.local/share/boundedcode/runtimes/llama.cpp/v0.5.0/bin/llama-bench
./bin/boundedcode doctor

Model weights are not distributed with this project. Each profile in configs/models/ names the upstream source, the pinned commit and the license. Download with scripts/fetch-model.sh (see the getting-started guide).

Limitations

• Pre-alpha. Commands and configuration may change between versions.
• Platform. Linux x86-64 with an NVIDIA GPU, tested only on the reference laptop below. Other platforms and GPUs are untested.
• CLI only. There is no daemon and no GUI.
• Verification presets exist only for Go and TypeScript. The TypeScript preset is limited: tsc, lint, test and build run only when the repository's node_modules and package.json scripts provide them. Other languages need a .boundedcode/verification.yaml.
• Cross-service analysis covers HTTP routes and calls, Kafka-style topics, environment variables and Terraform topic/queue resources, in Go and TypeScript/JavaScript. gRPC, OpenAPI, protobuf and SQL contracts are not analyzed (see cross-service-analysis.md).
• Frontier escalation works only through the Codex CLI with a ChatGPT subscription sign-in, or by pasting packets manually. API keys are refused (ADR-0009).
• Local-only success rates are measured only on this repository's benchmark suite of 11 synthetic tasks. They say nothing yet about real codebases.

Reference hardware

The project is developed and benchmarked on a laptop with an i7-13620H, an RTX 4060 Laptop GPU (8 GB), 64 GB DDR5 RAM and Debian 13. It is designed for this class of machine, not for datacenter GPUs.

License

Apache-2.0 (LICENSE). Third-party components and their licenses are listed in THIRD_PARTY_NOTICES.md and docs/licensing/upstream-license-matrix.md. Contributions use the DCO; see CONTRIBUTING.md.

This project is independent and is not affiliated with, sponsored by, or endorsed by the upstream projects or vendors it integrates with.