Show HN: AgentBridge – Let one AI think while another AI writes the code

AgentBridge, a new open-source tool released on GitHub and crates.io, lets developers split AI coding work by using one AI for reasoning and another for code execution, connected via the Model Context Protocol (MCP). The tool, which currently uses OpenCode as its local executor, aims to reduce subscription quota usage by separating planning from implementation, with a workflow of PLAN → EXECUTE → REVIEW → DONE.

Use one AI to think and another AI to code. AgentBridge connects AI reasoning environments to local coding agents through MCP , allowing a web-based AI to understand and plan changes while a local coding agent executes them in your workspace. Let the AI with the best reasoning access think. Let the coding agent you already use execute. 🇨🇳 中文 /IndexFlowing/AgentBridge/blob/main/README zh.md · 📦 crates.io https://crates.io/crates/agentbridge · 🐙 GitHub https://github.com/IndexFlowing/AgentBridge AI coding tools increasingly combine two different jobs: Reasoning — understanding a codebase, investigating problems, designing solutions, and reviewing changes. Execution — editing files, running commands, running tests, and applying changes. These two jobs do not necessarily need to be performed by the same AI. You may have a web AI with generous usage and strong reasoning capabilities, while your coding CLI has a more limited subscription quota. Without a bridge, the coding agent has to spend its quota on everything: Understand → Explore → Reason → Plan → Code → Test → Review AgentBridge separates the workflow: BRAIN Web-based AI Gemini / Claude / ... │ Reason / Plan │ MCP ▼ ┌─────────────────┐ │ AgentBridge │ └────────┬────────┘ │ C2C PLAN │ ▼ EXECUTOR Local coding agent OpenCode │ Edit / Run / Test │ ▼ Local Workspace The result is simple: Use the AI that is best at thinking, and the coding agent that is best at doing. AgentBridge introduces two explicit roles. The Brain is the AI responsible for reasoning. It can: - inspect the project; - search the codebase; - understand architecture; - investigate bugs; - design implementation strategies; - create implementation plans; - inspect Git diffs; - review the Executor's work. The Brain interacts with the workspace through AgentBridge's read-only MCP interface . It does not directly modify files or execute shell commands. The Executor is the local coding agent responsible for execution. It can: - modify files; - run commands; - run tests; - implement the Brain's plan; - report execution results. AgentBridge currently uses OpenCode as its Executor. The architecture is designed so additional coding agents can be supported in the future. AgentBridge connects the two. It provides: - MCP-based workspace access; - read-only project inspection; - structured task delegation; - task lifecycle management; - execution status; - Git diff inspection; - test status; - result reporting. A typical workflow looks like this: User │ ▼ Brain │ ├── Inspect workspace ├── Understand architecture ├── Investigate problem └── Create PLAN │ ▼ AgentBridge │ C2C PLAN │ ▼ Executor │ ├── Edit files ├── Run commands └── Run tests │ ▼ Git Diff / Result │ ▼ Brain │ Review │ ┌────┴────┐ │ │ DONE PLAN AGAIN This creates a feedback loop: PLAN → EXECUTE → REVIEW → DONE ↑ │ └── PLAN ────┘ The Brain can therefore focus on high-value reasoning while the Executor focuses on actually changing the codebase. AI coding subscriptions are often metered differently from normal web or chat usage. A coding agent may consume its allowance while: - exploring the repository; - reading files; - searching for definitions; - understanding architecture; - reasoning about an implementation; - generating a plan; - implementing changes; - running tests; - retrying failed implementations. That means a significant amount of coding-agent usage can happen before the first useful code change . AgentBridge lets you move much of the exploratory and reasoning-heavy work to another AI interface. For example: Gemini Web │ │ understand / reason / plan ▼ AgentBridge │ │ compact implementation plan ▼ OpenCode CLI │ │ implement / test ▼ Your repository The goal is not to bypass quotas. AgentBridge: - does not provide additional model credits; - does not bypass provider limits; - does not access paid models without authorization; - does not proxy model APIs. It simply allows you to use the AI services and coding agents you already have more efficiently . AgentBridge uses the Model Context Protocol MCP to expose your local project to the Brain. The Brain can use tools such as: | Tool | Purpose | |---|---| workspace info | Inspect workspace information and Git state | list directory | Explore the project structure | read file | Read files | search workspace | Search source code | git status | Inspect repository status | git diff | Review changes | test status | Read the latest test result | execution summary | Read the latest Executor result | The Brain does not receive a generic shell interface. It also cannot directly write files. This separation is intentional. AgentBridge uses a small structured protocol called C2C Context-to-Context to communicate between the Brain and Executor. Instead of passing an entire repository or a huge conversation to the coding agent, the Brain creates a compact implementation plan: C2C STATE: PLAN TASK ID: c2c 12345 ITERATION: 1 GOAL: Add URL inspection support to the GSC client. ACTIONS: 1. Inspect the existing GSC client. 2. Add URL inspection support. 3. Add tests for indexed and non-indexed URLs. TESTS: cargo test SUCCESS CRITERIA: Tests pass and the API correctly reports indexed / non-indexed. The source code remains in the local workspace. Only the task intent and implementation contract cross the Brain → Executor boundary. This keeps the communication focused and avoids unnecessarily duplicating the entire project context. One of the most important design decisions in AgentBridge is the trust boundary. The Brain can inspect: workspace ├── source files ├── project structure ├── Git status ├── Git diff └── test results But it cannot: ✗ write files ✗ delete files ✗ execute shell commands ✗ commit ✗ push The Executor is the component that performs those actions. Read-only │ ▼ ┌───────────┐ │ Brain │ └─────┬─────┘ │ PLAN │ ▼ ┌───────────┐ │ Executor │ └─────┬─────┘ │ Write / Execute │ ▼ Local Project The Brain decides what should happen . The Executor performs the implementation inside the local coding environment . The easiest way to install AgentBridge is through Cargo: cargo install agentbridge Verify the installation: agentbridge --version You can also download pre-built binaries from GitHub Releases https://github.com/IndexFlowing/AgentBridge/releases . git clone https://github.com/IndexFlowing/AgentBridge.git cd AgentBridge cargo install --path . Start the MCP server: agentbridge serve By default, AgentBridge listens on: http://127.0.0.1:8787/mcp The default configuration is intentionally localhost-only. You can check your environment with: agentbridge doctor Connect an MCP-capable AI client to AgentBridge. Your Brain can then inspect the local workspace through the MCP tools. The Brain instructions are provided in: skill/SKILL.md The skill teaches the Brain how to: - inspect the workspace; - understand the task; - create a C2C PLAN; - delegate the task; - monitor execution; - inspect the result; - review the changes; - finish or create another iteration. For example: Add Google Search Console URL inspection support to this project. First understand the existing architecture. Then create an implementation plan and delegate it to the Executor. After implementation, review the diff and test results. The Brain can inspect the actual repository instead of relying on files pasted into the conversation. If your Brain runs in a web environment and cannot directly access localhost, you can expose AgentBridge through a tunnel. For example, with Cloudflare Tunnel: agentbridge serve --allow-any-host Then: cloudflared tunnel --url http://127.0.0.1:8787 Your MCP endpoint will be available at: https://<your-tunnel-id .trycloudflare.com/mcp AgentBridge itself remains a local application. Security:If you expose AgentBridge outside localhost, use authentication and carefully choose which workspace is exposed. Do not point AgentBridge at your entire home directory. agentbridge init <workspace agentbridge serve agentbridge status agentbridge doctor agentbridge task start --goal "..." agentbridge task executed \ --status success \ --tests "cargo test" \ --exit-code 0 Run: agentbridge doctor to check your local AgentBridge environment. Global configuration: ~/.agentbridge/config.toml Workspace-specific configuration: <workspace /.agentbridge.toml Example: workspace = "/absolute/path/to/project" host = "127.0.0.1" port = 8787 security max file size = 1048576 deny sensitive files = true AgentBridge is designed around a simple security model: The Brain receives a read-only view of the workspace you explicitly expose. Path traversal and sensitive files are restricted. Examples of protected paths and patterns include: ../ /etc/passwd C:\Users\... ~/.ssh .env .pem .key id rsa Recommended practices: - Point AgentBridge at a single project. - Do not expose your home directory. - Keep the default localhost binding whenever possible. - If you expose the server remotely, configure authentication. - Treat a remote Brain as an external service with access to the workspace you expose. AgentBridge is a local developer tool, not a multi-tenant security boundary. AgentBridge is intentionally built around clear boundaries: AgentBridge │ ├── MCP Server │ └── Exposes workspace inspection tools │ ├── Workspace │ └── Secure filesystem access │ ├── Task Runtime │ └── PLAN → EXECUTE → RESULT lifecycle │ ├── C2C Protocol │ └── Brain → Executor communication │ ├── Executor │ └── Runs the local coding agent │ └── Git / State └── Tracks changes and execution results The key boundary is: Remote Brain │ MCP API │ ┌───────▼───────┐ │ AgentBridge │ └───────┬───────┘ │ C2C PLAN │ ┌───────▼───────┐ │ Executor │ └───────┬───────┘ │ Local process │ ┌───────▼───────┐ │ Workspace │ └───────────────┘ AgentBridge is not another AI coding agent . It does not try to replace: - Gemini - ChatGPT - Claude - OpenCode - Codex - your editor - your existing development workflow Instead, it connects them. AgentBridge does not: - provide AI models; - provide model credits; - bypass subscription limits; - proxy model APIs; - upload your repository to a hosted service. It is a local bridge between AI reasoning and local code execution . AgentBridge is currently focused on the Brain / Executor workflow. Current capabilities include: - Rust-based local MCP server - workspace inspection - secure path handling - Git status and diff inspection - structured C2C task protocol - task lifecycle management - OpenCode Executor integration - execution status and result reporting - Brain skill instructions - local-first architecture The Executor abstraction is designed to support additional coding agents as the project evolves. Potential future directions include: - Additional Executor backends - Executor selection and routing - Better task orchestration - Parallel task execution - Context optimization - Persistent task history - More Brain integrations - IDE integration - Richer review workflows The goal is not to build another monolithic AI coding product. The goal is to make the AI coding stack composable . Requirements: Rust 1.88+ Git Run: cargo fmt --check cargo clippy --all-targets --all-features -- -D warnings cargo test cargo build --release AI coding does not have to be a single-agent problem. Different AI products have different strengths, interfaces, context windows, pricing models, and usage limits. Instead of forcing one agent to handle everything, AgentBridge treats AI coding as a distributed workflow: THINK │ ▼ PLAN │ ▼ EXECUTE │ ▼ REVIEW │ ▼ DONE Use the AI that is best at thinking. Use the coding agent that is best at doing. Use AgentBridge to connect them. Contributions are welcome. If you want to add a new Executor, improve the MCP interface, or enhance the Brain / Executor workflow, feel free to open an issue or pull request. MIT License.