forgeplane/pinrail

Human in the loop for agentic workflows

Rust

1

1,040 commits

updated Oct 7, 2026

See the code

See what people are saying

README

Pinrail

Website · Download · Documentation · Plugins

Pinrail is a desktop app where your coding agents ask you before they act. You review what an agent proposes in a view made for it, and the agent carries on with your decision.

Pinrail with a code review open: the agent's findings on the diff, one accepted with a note to the agent and one rejected with a reason, and the other waiting reviews in the sidebar

Pinrail works with any agent that can run a command, such as Claude Code, Codex, Cursor or OpenCode. It runs on macOS and Linux, and Windows support is coming. Pinrail is in early development, so expect rough edges and changes between releases.

Why

Agents are good at doing the work, but some steps need a person: a comment posted under your name, an email to a customer, a change to production. Approving those steps in a chat means reading a wall of text and typing your answer back. Pinrail gives each of these moments a proper review:

  • The agent decides when to ask. Its instructions name the steps that need you, so it asks at those steps and nowhere else.
  • Each review has a view made for its content. A code review shows the diff with the agent's findings on it. A set of generated images shows the images, and you draw a box on the part that should change.
  • Your decision is structured. The agent gets back what you accepted, what you rejected and your notes, as Markdown it can act on or JSON for a script.
  • Everything stays on your machine. The app serves its API on loopback only, and keeps your reviews and decisions locally.

How it works

sequenceDiagram
  participant A as Your agent
  participant P as Pinrail
  A->>P: pinrail submit code-review --data review.json --wait
  Note over A: waits
  Note over P: you review and decide
  P->>A: your decision
  Note over A: carries on with it

When the agent reaches a step that needs you, it runs the pinrail command with a review: the plugin to show it with, a title and a JSON payload.

pinrail submit code-review --title "Retry failed webhook deliveries" \
  --data findings.json --wait

Pinrail notifies you, and the review waits in your inbox, grouped by project.

The inbox: reviews from several agents and projects, each with what it asks

You open it, and decide in the plugin's view. In the code review at the top of this page, you accept or reject each of the agent's findings, and write a note to the agent.

When you hand the review over, the command prints your decision and exits, and the agent carries on:

r_01K5R2 · decided · Retry failed webhook deliveries
code-review · decided by maya at 2026-09-23 10:14

- **#1 accepted** `src/deliver.ts:42` — The worker sleeps for up to 31 seconds per delivery (major)
  > Agreed. Re-enqueue with runAt = now + backoff(attempt)
- **#4 rejected** `src/log.ts:18` — The give-up log should say why (nit)
  > Fine as it is; the log already has the delivery id.

Undecided: #2, #3, #5

The command's exit code says how the review ended: decided, discarded with an instruction to stop, withdrawn or expired. When you ask for changes, the agent submits a new round, and the app shows your earlier verdicts beside it.

Getting started

  1. Install the app from the download page or the latest release. On macOS, it is signed and notarized.

  2. Follow the setup that opens the first time. It installs the pinrail command, adds Pinrail's skill to the agents it finds on your computer, and installs the recommended plugins.

  3. Ask your agent for something, in your own words:

    Ask me through Pinrail which TODOs in this repository to tackle first.
    

    Or ask it where Pinrail would help in your project:

    Look at this project and suggest where you should ask me through Pinrail before you act.
    

Your first review walks through this, and Instructing an agent shows how to make an agent ask at the steps you choose, every time.

Plugins

Each kind of review is a plugin: the payload an agent sends, the decision you give back, and the view you decide in. Five core plugins come with the app:

PluginReview
code-reviewA diff with the agent's proposed review comments, each accepted, rejected or edited.
listItems grouped under headings, each accepted or rejected with a note.
feedbackQuestions answered in one pass: choices, yes or no, and free text.
markdownA document, such as a plan or a spec, read and commented on section by section.
imageGenerated images to choose between, with boxes and pins on what to change.
The image plugin: candidate illustrations, with a box drawn on the part to changeThe feedback plugin: questions with choices, yes or no, and a comment

Sample plugins in forgeplane/pinrail-plugins show what else a plugin can do. They cover HTML pages, emails, calendars, logos, colour palettes, 3D models, animations, audio, video, design canvases, before-and-after comparisons and trades. Each one is released as a zip, which you install in the app's Settings › Plugins.

The sample plugins in forgeplane/pinrail-plugins, each showing a review

Writing a plugin

A plugin is a manifest, two JSON schemas and an HTML view, with no build step needed. pinrail plugins new <name> creates one, and the plugin SDK, pinrail-sdk, runs its view in a browser without the app and tests it with Playwright. Views can also be built with React, Vue, Svelte or any other framework. See Writing a plugin.

Repository

DirectoryContents
desktop/The app: a Rust core (API, storage, plugins), a Tauri shell and a React UI.
cli/The pinrail command that agents run.
plugins/The core plugins.
sdk/The plugin SDK, its development shell and its test harness.
skill/The skill that teaches an agent to use Pinrail.
e2e/End-to-end tests of the CLI and the app's UI against the headless core.
docs/The documentation, published on the website.
website/The website.

Development

Node is pinned in mise.toml, which mise install sets up, and Rust in rust-toolchain.toml, which rustup reads on its own.

mise run dev:desktop        # the app with live reload
mise run test               # every test suite
mise run lint               # rustfmt, clippy, the type check and the licences, as CI runs them

CONTRIBUTING.md describes each test suite and how to run it. Read it before you open a pull request. To report a vulnerability, follow SECURITY.md.

License

Pinrail is licensed under the Apache License 2.0. See NOTICE for the copyright and the third-party notices.

agentic-ai
agentic-workflow
agents

forgeplane/pinrail

Human in the loop for agentic workflows

Rust

1

1,040 commits

updated Oct 7, 2026

See the code

See what people are saying

README

Pinrail

Website · Download · Documentation · Plugins

Pinrail is a desktop app where your coding agents ask you before they act. You review what an agent proposes in a view made for it, and the agent carries on with your decision.

Pinrail with a code review open: the agent's findings on the diff, one accepted with a note to the agent and one rejected with a reason, and the other waiting reviews in the sidebar

Pinrail works with any agent that can run a command, such as Claude Code, Codex, Cursor or OpenCode. It runs on macOS and Linux, and Windows support is coming. Pinrail is in early development, so expect rough edges and changes between releases.

Why

Agents are good at doing the work, but some steps need a person: a comment posted under your name, an email to a customer, a change to production. Approving those steps in a chat means reading a wall of text and typing your answer back. Pinrail gives each of these moments a proper review:

  • The agent decides when to ask. Its instructions name the steps that need you, so it asks at those steps and nowhere else.
  • Each review has a view made for its content. A code review shows the diff with the agent's findings on it. A set of generated images shows the images, and you draw a box on the part that should change.
  • Your decision is structured. The agent gets back what you accepted, what you rejected and your notes, as Markdown it can act on or JSON for a script.
  • Everything stays on your machine. The app serves its API on loopback only, and keeps your reviews and decisions locally.

How it works

sequenceDiagram
  participant A as Your agent
  participant P as Pinrail
  A->>P: pinrail submit code-review --data review.json --wait
  Note over A: waits
  Note over P: you review and decide
  P->>A: your decision
  Note over A: carries on with it

When the agent reaches a step that needs you, it runs the pinrail command with a review: the plugin to show it with, a title and a JSON payload.

pinrail submit code-review --title "Retry failed webhook deliveries" \
  --data findings.json --wait

Pinrail notifies you, and the review waits in your inbox, grouped by project.

The inbox: reviews from several agents and projects, each with what it asks

You open it, and decide in the plugin's view. In the code review at the top of this page, you accept or reject each of the agent's findings, and write a note to the agent.

When you hand the review over, the command prints your decision and exits, and the agent carries on:

r_01K5R2 · decided · Retry failed webhook deliveries
code-review · decided by maya at 2026-09-23 10:14

- **#1 accepted** `src/deliver.ts:42` — The worker sleeps for up to 31 seconds per delivery (major)
  > Agreed. Re-enqueue with runAt = now + backoff(attempt)
- **#4 rejected** `src/log.ts:18` — The give-up log should say why (nit)
  > Fine as it is; the log already has the delivery id.

Undecided: #2, #3, #5

The command's exit code says how the review ended: decided, discarded with an instruction to stop, withdrawn or expired. When you ask for changes, the agent submits a new round, and the app shows your earlier verdicts beside it.

Getting started

  1. Install the app from the download page or the latest release. On macOS, it is signed and notarized.

  2. Follow the setup that opens the first time. It installs the pinrail command, adds Pinrail's skill to the agents it finds on your computer, and installs the recommended plugins.

  3. Ask your agent for something, in your own words:

    Ask me through Pinrail which TODOs in this repository to tackle first.
    

    Or ask it where Pinrail would help in your project:

    Look at this project and suggest where you should ask me through Pinrail before you act.
    

Your first review walks through this, and Instructing an agent shows how to make an agent ask at the steps you choose, every time.

Plugins

Each kind of review is a plugin: the payload an agent sends, the decision you give back, and the view you decide in. Five core plugins come with the app:

PluginReview
code-reviewA diff with the agent's proposed review comments, each accepted, rejected or edited.
listItems grouped under headings, each accepted or rejected with a note.
feedbackQuestions answered in one pass: choices, yes or no, and free text.
markdownA document, such as a plan or a spec, read and commented on section by section.
imageGenerated images to choose between, with boxes and pins on what to change.
The image plugin: candidate illustrations, with a box drawn on the part to changeThe feedback plugin: questions with choices, yes or no, and a comment

Sample plugins in forgeplane/pinrail-plugins show what else a plugin can do. They cover HTML pages, emails, calendars, logos, colour palettes, 3D models, animations, audio, video, design canvases, before-and-after comparisons and trades. Each one is released as a zip, which you install in the app's Settings › Plugins.

The sample plugins in forgeplane/pinrail-plugins, each showing a review

Writing a plugin

A plugin is a manifest, two JSON schemas and an HTML view, with no build step needed. pinrail plugins new <name> creates one, and the plugin SDK, pinrail-sdk, runs its view in a browser without the app and tests it with Playwright. Views can also be built with React, Vue, Svelte or any other framework. See Writing a plugin.

Repository

DirectoryContents
desktop/The app: a Rust core (API, storage, plugins), a Tauri shell and a React UI.
cli/The pinrail command that agents run.
plugins/The core plugins.
sdk/The plugin SDK, its development shell and its test harness.
skill/The skill that teaches an agent to use Pinrail.
e2e/End-to-end tests of the CLI and the app's UI against the headless core.
docs/The documentation, published on the website.
website/The website.

Development

Node is pinned in mise.toml, which mise install sets up, and Rust in rust-toolchain.toml, which rustup reads on its own.

mise run dev:desktop        # the app with live reload
mise run test               # every test suite
mise run lint               # rustfmt, clippy, the type check and the licences, as CI runs them

CONTRIBUTING.md describes each test suite and how to run it. Read it before you open a pull request. To report a vulnerability, follow SECURITY.md.

License

Pinrail is licensed under the Apache License 2.0. See NOTICE for the copyright and the third-party notices.

agentic-ai
agentic-workflow
agents