Protect the codebase
Every task runs in its own Git worktree. Your working branch changes only when a person runs the integration step.
Risk · no silent writes to your branchMeshloop · Loop engineering runtime
Let agents work in parallel. Keep your branch clean.
Meshloop turns an objective into a task graph you review, runs each task in its own Git worktree, and lets real compilers and tests decide what passes. Nothing reaches your branch until a person accepts it.
cargo install meshloop-cli --locked- Apache-2.0
- Windows · Linux
- One binary, no daemon
- CLI · slash commands · MCP
For engineering leadership
Parallel agents are a governance problem before they are a productivity win.
Multi-agent coding fails in predictable ways: work lands on the active branch, context costs grow with every file, repair loops burn budget, and models report their own success. Meshloop puts a deterministic boundary around each of those failure modes.
Contain the spend
Agents receive code skeletons (signatures, types, and docs) instead of whole files. Repair loops that stop improving are rolled back rather than retried indefinitely.
Cost · 67.7% fewer context tokens in the regression gateTrust the evidence
Your own compilers and test suites decide what passes, through real exit codes. A model never grades its own work.
Assurance · verification by ground truthIn one sentence
Meshloop lets a team add agent capacity while keeping codebase integrity, spend, and verifiable quality under explicit control — the three things leadership remains accountable for.
For adopters
Your first closed loop, in five steps.
Meshloop orchestrates the AI command-line tools you already use and have authenticated. It adds a plan you can review, isolation, verification, and an explicit hand-back.
Install one binary
A native Rust binary for Windows 10/11 and Linux. It starts on demand and exits cleanly, with no background service.
cargo install meshloop-cli --locked meshloop --version
Bundle the operator pack
From your repository root, write version-locked skills for Claude Code, Codex, Agy, Grok, and Pi, plus MCP tool definitions for MCP clients.
meshloop bundle --dest .
Point it at your harness
Add a minimal
meshloop.toml. Meshloop relies on your existing CLI sign-in and never writes credentials to disk.selected_harnesses = ["codex"] [limits] max_concurrent_workers = 2 [verify] verify_command = ["cargo", "check"]
Plan, review, run
Check the environment, decompose the objective into a task graph, then accept, decline, or adjust it before anything executes.
meshloop doctor meshloop plan "Add strict email validation" meshloop review-plan meshloop run
Integrate what you accept
Inspect the verified diff and evidence, then bring accepted work into your branch deliberately.
meshloop status meshloop integrate
Every capability is available three ways with the same behaviour: terminal commands, slash commands inside your agent, and MCP tools. Full guide: docs/start.md ↗
For technologists
Deterministic boundaries around non-deterministic agents.
Meshloop reuses proven foundations such as Git worktrees, SQLite WAL, and kernel process primitives, and concentrates new engineering on loop control: context reduction, convergence, and contention.
- Worktree isolation and replay
- Each task executes in
.meshloop-worktrees/<task-id>. State transitions are atomic SQLite WAL transactions, so interrupted runs replay from a consistent record. - Context reduction without a vector database
- AST skeletons keep signatures, types, and docs. Symbol ranking uses a 64-dimension Walsh-Hadamard transform with 2-bit quantization in pure Rust.
- No orphan processes
- Windows Job Objects (
KILL_ON_JOB_CLOSE) and POSIX process groups tie every compiler and test process to its task. - Convergent self-repair
- Diagnostics form a partially ordered lattice. A repair round must reduce the error potential; regressions reset hard and cycles terminate.
- Hexagonal crates, no unsafe code
- Domain, context, engine, adapters, and CLI crates. Core crates use
#![forbid(unsafe_code)]; the engine polls without an async runtime. - Contention-safe Git
- A Git admin mutex with exponential backoff resolves rival
.git/index.lockcontention between concurrent workers.
28/28regression-gate metrics passing
0orphan processes and leaked worktrees after runs
324 µsp95 quantized symbol search (target ≤ 500 µs)
1.48 sfull end-to-end smoke lifecycle (target ≤ 5 s)
Figures are reported by the repository's own regression gate for 0.1.0 (benches/thresholds.toml ↗). Reproduce them with cargo run --release -p xtask -- bench.
Status and boundaries
What you can rely on today.
An honest picture for teams deciding whether and where to try Meshloop.
| Release | 0.1.0, the first public release, published on crates.io as meshloop-cli. Start on a non-critical repository. |
|---|---|
| Platforms | Native Windows 10/11 and native Linux. It does not operate across the Windows/WSL2 file boundary. |
| Agent tools | Bundled skills for Claude Code, Codex, Agy, Grok, and Pi; MCP tool definitions for MCP-capable clients. |
| Deliberately excluded | No daemon or background service, no stored credentials, no external vector database, and no automatic merge into your branch. |
| Licence | Apache-2.0 for the code. The Meshloop name and mark are covered separately in TRADEMARKS.md ↗. |
Questions
Before you adopt.
Short answers for engineers, security reviewers, and budget holders.
Does Meshloop replace my coding agent?
No. It orchestrates the agent command-line tools you already use. Meshloop owns planning structure, isolation, verification, and hand-back; the agent still writes code.
Can it merge changes on its own?
No. Plans pass a human review gate before execution, and integration into your branch is a separate explicit step.
Does it send my code anywhere?
Meshloop runs locally and has no service of its own. Your configured AI tool communicates with its provider exactly as it would without Meshloop, but receives pruned context rather than whole files.
Do I need Docker, a server, or a vector database?
No. It is a single native binary that starts on demand and exits when the work is done.
How do I know the benchmark numbers are real?
They come from a regression gate in the repository with published thresholds. You can run the same gate on your machine.
Give your agents a runtime with boundaries.
Install Meshloop, run one closed loop on a small objective, and inspect the evidence before you integrate anything.
Built by Samuel Mota · Move the Needle
Why Meshloop exists
Open source · Apache-2.0 · 0.1.0 on crates.io
Value
Run AI coding agents in parallel without losing control of the codebase: reviewed task graphs, isolated worktrees, compiler-verified repair, and a human accept gate.
Strategy
Scaling agents is a governance problem before it is a productivity win. Meshloop keeps codebase integrity, spend, and verification under explicit control while the agent tools themselves stay interchangeable.
Technical frontier
Ephemeral Git worktrees per task, AST skeleton pruning across seven languages, kernel-owned process trees, and Lyapunov-bounded self-repair judged by real compiler exit codes, in one daemonless Rust binary.