tap protocol lets Claude and Codex collaborate on one codebase without shared runtime
Minseo Kim presents tap, a file-based collaboration protocol that enables Claude (Anthropic) and Codex (OpenAI) to jointly develop and review code on a shared codebase without requiring shared memory, a central server, or a common runtime.
Score breakdown
Heterogeneous model pairs using tap recorded defects or requested changes in 69.8% of reviews versus 53.1% for homogeneous pairs, demonstrating that cross-vendor agent collaboration on a shared codebase produces broader code review coverage than single-vendor setups.
- 01tap is a file-based protocol enabling Claude (Anthropic) and Codex (OpenAI) to collaborate on a shared codebase without shared memory or a common runtime.
- 02The protocol uses a two-tier design: Tier 1 is file-based message preservation (markdown files with metadata); Tier 2 adds real-time notification paths for each agent.
- 03Work isolation is achieved through separate git worktrees, and messages remain accessible even if real-time delivery fails or an agent restarts.
Minseo Kim presents tap, a file-based collaboration protocol that addresses a gap in multi-agent software development: existing systems typically assume a shared runtime, a central conversation server, or agents from the same API family, making cross-vendor agent collaboration difficult. tap solves this by treating markdown files with metadata as the primary, durable message format, allowing agents to communicate asynchronously across different execution environments.
The protocol operates on two tiers: Tier 1 is a file inspection path where messages are preserved as files, and Tier 2 adds real-time notification paths tailored for Claude and Codex.
The protocol operates on two tiers: Tier 1 is a file inspection path where messages are preserved as files, and Tier 2 adds real-time notification paths tailored for Claude and Codex. If real-time notification fails or an agent restarts, the underlying message file remains intact and can be re-inspected. Work isolation is achieved through separate git worktrees, preventing agents from interfering with each other's in-progress changes on the shared codebase.
To validate tap, the authors ran a 27-day, 37-generation self-applied operation in which tap was used to develop and review itself, producing 209 tap-related pull requests and 717 operational artifacts. Analysis of 375 review artifacts revealed that heterogeneous model pairs (Claude + Codex) recorded at least one defect or requested change in 69.8% of reviews, versus 53.1% for homogeneous pairs — suggesting that pairing agents from different vendors broadens review perspectives. tap is released as the open-source npm package `@hua-labs/tap` at version `v0.5.2`.
Key facts
- 01tap is a file-based protocol enabling Claude (Anthropic) and Codex (OpenAI) to collaborate on a shared codebase without shared memory or a common runtime.
- 02The protocol uses a two-tier design: Tier 1 is file-based message preservation (markdown files with metadata); Tier 2 adds real-time notification paths for each agent.
- 03Work isolation is achieved through separate git worktrees, and messages remain accessible even if real-time delivery fails or an agent restarts.
- 04A 27-day, 37-generation self-applied operation used tap to develop and review itself, generating 209 pull requests and 717 operational artifacts.
- 05Analysis of 375 review artifacts found heterogeneous model pairs recorded a defect or requested change in 69.8% of reviews, vs. 53.1% for homogeneous pairs.
- 06tap is distributed as the open-source npm package `@hua-labs/tap` at version `v0.5.2`.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 15, 2026 · 11:57 UTC. How this works →