Project mesh for AI coding runtimes

Keep every coding agent in sync.

AgentMesh gives each repository one local source of truth for instructions, skills, and subagents, then keeps Claude Code and Codex aligned through an auditable lockfile.

Install AgentMesh Quickstart

mesh previewbidirectional

One local meshChanges reconcile across connected runtimes before AgentMesh writes.

Any sourceedit or add anywhere

All viewsClaude, Codex, .ai/

Auditablelockfile plan

Install

Verified installers for macOS, Linux, and Windows.

Release installs download the platform archive, verify it against SHA256SUMS , verify the cosign signature bundle, and then place the single agentmesh binary.

macOS and Linux

Use the shell installer for local development environments and CI images.

curl -fsSL https://agentmesh.sh/install.sh | sh

Windows

Use the PowerShell installer for native Windows workstations and terminals.

irm https://agentmesh.sh/install.ps1 | iex

How it works

Edit anywhere. Sync everywhere.

AgentMesh keeps every runtime view and the project model aligned, no matter where instructions, skills, or subagents are added or changed.

AgentMesh project model.ai/

instructionsskillssubagents

Runtime view.claude/

CLAUDE.md.claude/skills.claude/agents

AgentMesh syncscan → normalize → merge → emit

Every pass records an auditable plan in agentmesh.lock.

Runtime view.codex/

AGENTS.md.codex/skills.codex/agents

Any sourceEdit or add in .claude/, .codex/, or .ai/.

Same modelInstructions, skills, and subagents reconcile together.

More runtimesCursor, Copilot, Gemini CLI, Windsurf, and others can join the mesh.

Core capabilities

Bidirectional sync without becoming an inference proxy.

AgentMesh does not run models, store prompts, or proxy API calls. It keeps repository context coherent across coding runtimes.

Canonical model

Instructions, skills, subagents

Runtime-native files are normalized into a shared entity model and tracked in agentmesh.lock.

Non-destructive

Conflicts stay recoverable

Losing versions are preserved in the machine-local cache and surfaced through doctor, restore, and ack.

Explicit trust

Hooks are installed by command

Runtime hooks are never installed silently. Install, upgrade, uninstall, and restore remain explicit user actions.

Adapters

Bundled adapters first, external adapters later.

Claude Code and Codex parsing and emitting run inside first-party Rust crates. External adapter trust tiers are reserved for later releases.

Read adapter docs

Bundled adapter

Claude Code

CLAUDE.md, .claude/skills, .claude/agents, local hook settings

Bundled adapter

Codex

AGENTS.md, .codex/skills, TOML agents, local hook settings

Security and migration

Explicit trust boundaries for files agents execute.

Hook records use absolute binary paths and machine-local integrity pins. Sync paths do not stop for prompts; ambiguous choices happen at init, install, restore, upgrade, or uninstall boundaries.

Distribution

Verified release artifacts

Archives are checked against signed checksum manifests before installation.

Releases

Versioned surfaces stay separate

The artifact, adapter protocol, lockfile schema, and entity schemas can evolve independently.