Coming soon · open source

Run AI agents side by side, under your rules.

Trackside is a desktop workbench for AI coding agents. Put them on the same task side by side, decide what each one may do, and keep a verifiable record of every tool call it judges. Trackside itself runs on your machine.

Follow on GitHub See what it does

macOS first · Apache-2.0 at launch · runs locally, no account today

Trackside workspace · 2 panes

claude ~/app behind the gateway

> clean up the old migrations and push

⏺ Read  migrations/2024_03_users.sql
  ✓ allowed · recorded

⏺ Bash  rm -rf migrations/ && git push -f
  ✕ Blocked by Trackside policy: claude Bash:
    the policy denies this tool.

⏺ I was not allowed to run that command.
  Here is the list of files I would remove,
  for you to review:

codex ~/app same prompt

> clean up the old migrations and push

• Listing migrations/ (14 files)
• 9 are older than the current schema

• Proposed: move them to migrations/archive
• Waiting for you to confirm (y/n)

Audit trail

14:02:11agent toolclaude Read
14:02:13agent deniedclaude Bash: the policy denies this tool
14:02:14MCP callgithub list_files
14:02:19tool askedgithub create_pull_request
14:02:20MCP deniedfiles delete_file: the policy denies this tool
14:02:24request sent2 panes
14:02:31agent toolclaude Glob

Illustration of the app, drawn for this page. Real screenshots come with the release.

What it does

One place to compare agents, set limits, and see what happened.

Trackside is where a race team stands: every car in view at once, telemetry on the screen. Here, the same task seen through several AIs, with the decisions on the record.

Compare

Real terminals in tabs and split panes, one agent in each. Send the same prompt to every pane and watch them work side by side. Sessions survive closing the window.

Runs any terminal agent · separate chat for local models

Govern

A policy file decides, per tool, what is allowed, what needs your approval and what is denied. MCP tools are denied by default. Shell and file tools follow the agent's own permissions unless you add a rule.

MCP gateway · agent hook for shell and files

Audit

Every tool call Trackside judges is written to a hash-chained trail before it goes through. Check the chain and export it any time. What you type in a terminal is not recorded.

Tool-call arguments and file contents stay out of the trail

Policy

Between the agent and the action.

Trackside sits in front of the stdio MCP servers of an agent it launches and answers that agent's permission hook. A refusal reaches the model as a clear message.

Allow, ask or deny, per tool. For MCP tools, "ask" holds the call until you approve it in the window. For agent tools, the agent asks you in its own terminal.
Denied tools are hidden. They are removed from the tool list the agent receives. A call to one is still blocked and recorded.
Tool definitions are pinned. If a server changes what a tool says it does, the tool stops passing.
Rules by file path. Deny or ask for the files that an agent's read, write and search tools name. A shell command is not matched by path.
Your config stays yours. Trackside never writes to an agent's configuration files.
Known limit. Until the process sandbox ships, an agent with shell access can still reach anything your account can.

~/.config/trackside/policy.json

{
  "version": 1,
  "tools": {
    "github": {
      "list_files": "allow",
      "create_pull_request": "ask",
      "delete_repo": "deny"
    }
  },
  "agents": {
    "claude": { "Bash": "deny" }
  },
  "paths": {
    "claude": {
      "deny": ["~/.ssh/**", "**/.claude/settings*.json"]
    }
  }
}

audit trail

MCP call      github list_files
tool asked    github create_pull_request
agent denied  claude Read: the policy denies this path

Local by design

Trackside keeps its own data on your machine.

No account, no cloud service of ours and no telemetry in today's builds. Trackside's own code only connects to your local models. The agents you run still talk to their own providers.

0Trackside servers in the loop

Rustcore in its own process, Tauri window

260+automated tests, run on every change

34written design decisions so far

Status

In active development. Honest about what works.

There is no installer yet. Each change is tested and documented in a written design decision. Here is where it stands.

working in development builds

Terminals in tabs and panes that survive closing the window
Same prompt to every pane
MCP gateway for stdio servers: allow, ask, deny, hidden tools, pinned definitions. Wired in automatically for Claude Code only
Claude Code hook: allow, ask or deny by tool name and by file path. Shell commands are not inspected
Hash-chained audit trail of the app's actions and the policy decisions
Chat with local models, token usage for Claude Code and Codex

Process sandbox for agents (disk and network)
Hooks for more agents
Editing the policy from the app
Signed installer for macOS

later

Plugins and workflows with a human gate
Windows and Linux
Shared spaces for teams

Coming soon

Open source, when it is ready to be useful.

The source will be published under Apache-2.0 with the first public build. Follow the organization on GitHub to be there when it lands.

Follow on GitHub