# orcareplay [Health: Active]

**Category:** 🤖 Coding Agents  
**Repository:** https://github.com/Continuum-AI-Corp/OrcaReplay  
**GitHub Stars:** 1059  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/orcareplay

## Description
Read, replay and fork recorded coding-agent runs.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "orcareplay": {
    "command": "uvx",
    "args": ["orcareplay-langgraph"]
  }
}
```

## Documentation

## What orcareplay MCP server does

orcareplay MCP server is designed for recording and examining coding-agent runs. It captures more than model traffic: the trace can include streamed output, tool calls and results, shell exit codes, timing, file writes, and other events observed by its capture layers. This gives developers a timeline for investigating what an agent did, including actions that may not be visible through an SDK wrapper.

A recorded run can be replayed from disk. Exact replay blocks unmatched model egress, so the model is not called and the replay does not consume additional model tokens. The project also supports starting from a selected checkpoint and continuing with another model. That makes it possible to compare model behavior while preserving the earlier conversation and workspace state.

## How it works

The CLI places a local proxy in front of the agent and supplies environment settings rather than requiring changes to the agent itself. The proxy observes requests and streamed responses, while additional capture layers cover process output, MCP configuration rewriting, workspace changes, and selected hardcoded network origins. Events are written into a single ordered trace.

During replay, the proxy serves matching responses from that trace. The default mode refuses requests that are not represented in the recording. A loose mode can forward an unmatched request to the provider and marks the replay as live on divergence. Recorded tool calls still execute for real, so replay is not a security sandbox; shell commands or MCP servers that open their own connections may access the network.

## Setup and configuration

The package is installed globally with npm and exposes the `orca` command:

```bash
npm i -g orcareplay
orca record claude
orca replay last
```

The repository states that Node.js 20 or newer is supported. An agent must be installed for the record and replay examples, and normal recording may require a provider key, network access, and model tokens. The `orca quickstart` command creates a small project, a genuine bug, and a sample recording so the workflow can be tried without supplying an existing agent setup.

## Tools and capabilities

The documented command-level capabilities include:

- Recording an unmodified coding-agent run.
- Replaying the latest or another saved run from its trace.
- Forking from a chosen step with a different model.
- Comparing outcomes by grading the resulting project, such as with `npx tsc --noEmit`.
- Capturing an agent’s assembled system prompt with the capture script.
- Inspecting runs after the original terminal has closed because traces are stored as files.

orcareplay MCP server also supports recording agents whose model endpoint cannot be redirected, using the optional TLS interception mode described by the project.

## Limitations and notes

It also does not provide client-specific setup for Cursor or other MCP hosts. Treat the documented interface as the `orca` CLI unless the repository’s current documentation supplies additional MCP configuration.

Exact replay only guarantees model-response reuse for requests present in the trace. Tools run again, and network behavior from those tools is outside that guarantee. TLS interception can record metadata for tunneled hosts, but it does not turn replay into isolation. Use a separate sandbox when filesystem or network side effects must be contained.

_Full upstream README: https://allmcps.com/mcp/orcareplay/readme_

