# Agent Interop Runtime [Health: Active]

**Category:** 🤖 Coding Agents  
**Repository:** https://github.com/Praket7/agent-interop-runtime  
**GitHub Stars:** 1  
**npm Downloads (last month):** 1221  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/agent-interop-runtime

## Description
A capability graded local MCP runtime for Freebuff, Codex, OpenCode, and Claude Code sessions.

## Tools
Capabilities this server exposes over MCP:

- **get_runtime_status** — Report installed provider capabilities and the active tool profile. Use list_providers for per-provider capabilities and list_agent_sessions to inspect available sessions.
- **list_providers** — List configured provider adapters and their supported actions. Use list_agent_sessions to find actual native sessions; discovery does not confirm sign in or readiness.
- **list_agent_sessions** — Find actual native sessions from installed providers, optionally filtered by provider. Provider errors are returned separately from an empty result. Use list_providers to inspect adapter capabilities.
- **get_session_diff** — Read file change evidence for one exact native provider session. Use read_session_events for its progress history; this tool returns file changes only.
- **read_session_events** — Read a bounded page of native provider events with a resumable cursor, stream epoch, and retention-gap metadata. Use this for provider event history; get_thread_progress reads Freebuff Desktop progress instead.
- **send_session_message** — Send one prompt to an exact native provider session. Use send_thread_message only for a Freebuff Desktop thread, or send_conversation_message when you need a durable shared transcript. Model, agent, and reasoning are separate controls. For OpenCode, accepted means queued by transport and does not mean completed.
- **cancel_session** — Cancel work in a provider native session.
- **create_session** — Create a native provider session when supported.
- **resume_session** — Resume or reattach to an exact native session.

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

```json
"mcpServers": {
  "agent-interop-runtime": {
    "command": "npx",
    "args": ["-y","agent-interop-runtime@1.1.0"]
  }
}
```

## Documentation

## What Agent Interop Runtime does

Agent Interop Runtime gives an MCP client a local coordination layer for coding-agent sessions. It can discover supported Freebuff, Codex, OpenCode, Claude Code, and Cursor sessions installed on the same computer, then expose the controls each provider supports.

The runtime keeps provider boundaries intact. Accounts, permissions, histories, project access, login flows, and consent prompts remain managed by the provider. The runtime records delivery status and provider updates, but a delivery receipt only confirms that a provider accepted work; completion requires an observed result.

When delivery becomes uncertain after a network failure, the runtime marks the request as unknown rather than replaying it. This avoids unintentionally submitting the same work twice.

## How it works

An MCP client starts the runtime, preferably as a local process over standard input and output. The runtime checks which provider sessions are available, reports their supported controls, and routes work to the selected session. Provider capability depends on the provider version, local configuration, and account state.

The runtime can create or resume sessions when a provider supports those operations. It can send tasks, read progress, wait for updates, request cancellation, and retrieve provider diffs where available. Shared work records, notes, handoffs, reviews, evidence, and coordination claims allow sessions using the runtime to exchange task context.

Optional verification runs local checks inside a named workspace using programs explicitly configured by the user. These commands are not sandboxed. File access is limited to small text files in an approved project folder; common credential files are blocked, credential-shaped text is redacted, and binary files are refused.

## Setup and configuration

Use Node.js 20 through 26. Install the current major release with npm or pnpm, then inspect available providers with the doctor command:

```sh
npm install --global agent-interop-runtime@1
agent-interop-runtime doctor
```

A source checkout can be built with pnpm and inspected through the generated CLI. For MCP connections, start `serve` through a client's local-server configuration. Claude Code can register the server with a stdio command, while Codex, Cursor, and OpenCode have setup or configuration examples in the project. Other MCP clients can launch the same command through their local server settings.

Tool exposure is selectable through profiles. Pass `--profile minimal`, `core`, `freebuff`, `full`, or `legacy` to `serve`, or set `INTEROP_TOOLS_PROFILE` when the client cannot pass arguments. The minimal profile exposes nine focused tools; broader profiles add coordination and Freebuff-specific operations. `INTEROP_READ_ONLY=1` removes tools that change provider or coordination state.

HTTP mode is intended for deliberate remote setups and requires authentication plus network protection. Session identifiers alone are not access control.

## Tools and capabilities

- Discover installed provider sessions and inspect their reported capabilities.
- Create or resume sessions where the provider supports those actions.
- Send work, monitor progress, wait for updates, request cancellation, and collect diffs where supported.
- Maintain shared work notes, handoffs, evidence, reviews, and resource claims.
- Run explicitly enabled local verification checks in a named workspace.
- Read approved small text files with credential filtering and binary-file rejection.
- Select a modern or legacy tool catalog based on client and migration needs.

The legacy profile preserves the original 0.3.x tool names, while modern profiles use consolidated, verb-first operations.

## Limitations and notes

Finding a provider does not confirm that it is signed in or ready to accept work. Provider applications must be installed on the computer running the runtime, and the runtime cannot bypass provider login, consent, or permission requests. It also does not replace the provider's own application or merge separate accounts and histories.

Resource claims coordinate agents that use this runtime but cannot prevent unrelated programs from editing the same files. Verification commands run with the access available to the user's account, so they should only be used with trusted work and commands. Automated checks and offline benchmarks use simulated providers rather than live sessions and do not measure speed or provider cost.

_Full upstream README: https://allmcps.com/mcp/agent-interop-runtime/readme_