gpt-agent-bridge

GPT-Agent Bridge

English | 简体中文

Think with ChatGPT Web. Coordinate your local Agent team.

GPT-Agent Bridge turns an ordinary ChatGPT Web Project Chat into the thinking partner and coordinator for work on your own computer — a Web Principal. You and ChatGPT frame the real problem and set the goal; ChatGPT dispatches bounded assignments to the local Agents you already run in Herdr; when an Agent finishes, ChatGPT fetches the available report and checks it against your actual files, diffs and images before the next step. You keep control of goals, scope and permissions; Bridge does not install providers or supply Agents.

For example: "investigate this failing test from two angles, then have a third Agent review the fix." While your Agents work you can keep reasoning in the Chat — or step away from the screen; when one finishes, an optional Chrome reminder can notify the bound Chat, where the Web Principal picks the work back up.

See it work

Demo · ChatGPT Web ↔ Bridge ↔ four Herdr Agents (2×, 60 fps, silent)

https://github.com/user-attachments/assets/f1f3fb48-466e-4154-8b79-6d5a587ffd16

One continuous recording: four assignments, local PDF analysis, report review and a worker error. Use fullscreen for terminal text.

One round of work

ChatGPT decides → your local Agents execute → results return to the Chat → ChatGPT checks the evidence → you continue.

1. Decide. You and ChatGPT sharpen the question, compare options and set scope and acceptance criteria.
2. Dispatch. ChatGPT sends bounded assignments to the eligible Herdr Agents you authorized — in parallel, across providers and models.
3. Return. When an Agent finishes, the optional browser extension can send a completion notice to the bound ChatGPT Chat. The Web Principal receives it, reads the available report, checks the results and decides the next step.
4. Verify. Before trusting any summary, ChatGPT reads the real project files, diffs and images behind it.
5. Continue. Convinced? Take the next step. Something off? Re-check or re-assign — every dispatch still needs your authorization.

What ChatGPT can see in your project

• Text and images, read directly. ChatGPT can read UTF-8 source files (long files page by page) and native PNG/JPEG/GIF/WebP images (up to 5 MiB each) — no copy-paste.
• The repository, inspected in place. Project context, directory listings, text search, diffs and file metadata (size, SHA-256) answer "what does the code actually look like" — read-only, inside the project you connected.
• Original files on demand. For PDF/DOCX/ZIP originals, ChatGPT asks Bridge to prepare the file and then fetches the exact bytes itself, in chunks with SHA-256 checks — no more re-uploading attachments by hand. Preparing is not importing: nothing is parsed until ChatGPT actually fetches it, and ordinary read never extracts PDF/DOCX/ZIP contents.

While you're away

After a local Agent finishes, the optional Chrome extension can automatically send a short completion notice (Chinese or English) to the one ChatGPT Chat you explicitly bound and enabled. The Web Principal in that Chat receives it, reads the report when one is available, checks the actual results and decides the next step — you don't have to keep watching the screen. This works only while Chrome with the extension, the local Bridge service and that Chat are actually available; it is not an always-on autopilot, and finishing is not the same as succeeding. It refuses to send while the page is busy; with a draft it offers append-on-a-new-line or ignore and never presses Send for you; notices are not duplicated, and an effect it cannot confirm is not retried.

Continue in a new chat

When the Chat's context grows long, tell ChatGPT: "Use the handoff Skill to sync this Chat's project memory — I'm switching to a new Chat." It saves what matters for the next judgment into the project's local notes — not a full transcript, and not every detail is guaranteed to survive. Then open a new Chat in the same ChatGPT Project and just say: "Continue." The Web Principal reads the saved state, rechecks the repository and what is actually running, and carries on — no copying chat history over. Saying "Continue" does not enlarge earlier authorizations, and old work is not revived; what it finds is rechecked on the spot.

Why Web Principal

Bridge is not a one-way "fire a command and hope" button, and not another separate coding dashboard: the Chat where you already think stays the place where goals are set, results are questioned and next steps are chosen.

• Examine the question before dispatch. Reframe unclear requests, expose missing premises and compare alternatives before choosing a plan.
• Coordinate multiple independent Agents. Give your existing Herdr Agents bounded questions in parallel, across providers and models. Separate implementation from review, and ask one Agent to challenge another's findings.
• Check evidence before accepting results. Read complete task reports and inspect actual code, files, diffs and test evidence. Distinguish observations from claims, gaps and unknown outcomes; agreement alone does not prove correctness.
• Carry decisions across chats deliberately. When authorized, preserve the decision, reasons and next boundary in project notes; the four installed Skills help a new Chat pick the thread back up. This does not restore running tasks, replay them or automatically create global memory.

For example, when three eligible Herdr Agents are already available, explore a storage-design decision with ChatGPT. Have two investigate independent questions in parallel, then ask the third to review and challenge them. Check the reports against the relevant code, diffs and test evidence before choosing an approach. Each assignment still needs your authorization.

For demanding reasoning and decisions, optionally use GPT-6 Pro as the Web Principal while Herdr-managed Agents implement and review. You select the model through your own eligible ChatGPT plan and workspace. Bridge does not include model access, select it automatically or promise comparative speed or quality.

Download: the complete installation ZIP gpt-agent-bridge-v0.1.0-rc.8.zip · release page.

Start here: the complete setup guide →

Your computer	Core preparation	Web and optional reminders
Native Ubuntu Desktop	Ubuntu Terminal: ./prepare	ChatGPT Web; optional Chrome reminders on the same Ubuntu computer
Windows + WSL2	Windows PowerShell: .\Start-Bridge.ps1 prepare	Same Web setup; optional Windows Chrome reminders; core runs in WSL2

Follow the guide from step 1 before using either command. You need existing Herdr 0.9.0 in the chosen Linux environment and the required ChatGPT workspace controls. Check custom MCP, Plugin Creator/Skills and Project Instructions access before local installation. You do not need a running Agent just to set up Bridge. Both routes continue with the same ChatGPT Web setup; reminders are optional.

How it works

1. Prepare your computer. The guide explains what you approve: a private Python environment and two Linux background services (Bridge and its Tunnel connection), installed, enabled and started. Chrome extension preparation and loading happen only if you later choose reminders.
2. Connect ChatGPT. A private Secure MCP Tunnel carries tool requests from ChatGPT to your Bridge. Install the connection as a plugin, then use Plugin Creator in ordinary Web Chat with the package's single Bridge-Skills.zip to add the four supplied Skills — reusable instructions for connecting, resuming, handing off and coordinating work.
3. Choose the project and permissions. Save the supplied base instructions (Core) in your ChatGPT Project. Connect an existing local Git repository, review the returned project-identifying text (Binding), and save it beside Core. If Bridge recognizes no usable continuity configuration, onboarding asks once whether to set up notes; saying no leaves them unwritten. Start with a small read-only check.

The guide closes with simple checks: one visible bridge_health call in a Project Chat and the four installed Skill names. Creating the connection does not start Agents or authorize project work.

Plugin Creator and installation depend on plan/workspace permissions. If the required controls are unavailable, stop at that account issue. Step 6 uses one visible Bridge-Skills.zip upload to ask Plugin Creator to update your existing personal Bridge plugin — preserving its MCP connection and name — then confirms the four Skill entries and a new ordinary Project Chat health call. No hidden folders or separate file uploads are involved. Work/local Codex is not part of this setup. No separate model API Principal is required.

The optional Chrome extension sends completion reminders only to the exact Chat you explicitly select. On Ubuntu, it runs on the same computer as Bridge; on Windows, load the final private Windows-local folder printed by the optional reminders action (./prepare reminders or .\Start-Bridge.ps1 reminders). It preserves drafts and does not send an appended reminder for you.

After setup: three useful steps

1. Orient: select the authorized project and ask for current context and a read-only check of a named file. Verify the repository. When you need delegation, start your supported Agent in Herdr and check that it is actually available.
2. Delegate deliberately: state the question, allowed changes and acceptance criteria; authorize input to an observed eligible Agent. Continue useful independent reasoning while it works.
3. Judge and preserve: read the full report, compare original evidence, resolve contradictions and save decision-relevant project notes when authorized. A reminder proves neither success nor completed review.

Accounts, provider sign-in, model selection, fees and quotas remain yours.

Download

Download the complete installation ZIP gpt-agent-bridge-v0.1.0-rc.8.zip from the v0.1.0-rc.8 release page. Use the full release ZIP; GitHub's Source code ZIP and Code → Download ZIP are not installation packages. The package for the setup steps must contain Bridge-Skills.zip at its root next to prepare, Start-Bridge.ps1 and release-manifest.json; confirm this after extracting.

Next: follow the beginner guide. Herdr prerequisite: the short installation section. Advanced manual setup and optional example: Quickstart. For service inspection and advanced manual stop/removal: autostart. Advanced contracts and documentation navigation: technical reference. Project notes: continuity. Task, reminder and restart limits: lifecycle.

MIT License.