Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@harness-homies what are my currently running coding-agent sessions?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
A read-only MCP server that lists coding-agent sessions on your machine — Claude Code, OpenCode, Cursor, and Codex — in one place: what's running, its topic, its todos, its transcript.
It never writes to another agent's session and never sends messages.
/plugin marketplace add nikp29/harness-homies
/plugin install harness-homies@harness-homiesThis installs the MCP server (run via npx -y harness-homies) and the inspect-agents skill.
plugin/ is an Agent Plugins 1.0.0 package: plugin.json, mcp.json, and skills/inspect-agents/SKILL.md. Point any conformant client at that directory. The same directory also carries .claude-plugin/plugin.json and .mcp.json for Claude Code.
OpenCode plugins are JS modules and can't register MCP servers or skills, so set it up with config instead:
Merge opencode/opencode.json into ~/.config/opencode/opencode.json (or a project's opencode.json).
Copy the skill: cp -r plugin/skills/inspect-agents ~/.config/opencode/skills/
codex mcp add harness-homies -- npx -y harness-homiesAdd an entry to ~/.cursor/mcp.json:
{ "mcpServers": { "harness-homies": { "type": "stdio", "command": "npx", "args": ["-y", "harness-homies"] } } }pnpm install
pnpm build
claude mcp add --scope user harness-homies -- node /path/to/harness-homies/dist/cli.jsRelated MCP server: Hua PlanRelay
Tool | Args | Does |
|
| List sessions across one or all agents, live and historical, most recent first |
|
| Title, cwd, status, todos, message count |
|
| Paginated transcript, secrets redacted by default |
|
| Substring search over titles and recent transcript content |
agent is one of claude-code | opencode | cursor | codex. All tools are readOnlyHint: true.
pnpm test # builds, then runs the suite against dist/
pnpm typecheck
node dist/cli.js doctor # per-adapter availability + session countsRedaction is on by default for transcript text (API keys, tokens, PEM blocks, etc.). Pass raw: true to skip it.
Claude Code: reads ~/.claude/sessions/*.json (live) and ~/.claude/projects/**/*.jsonl (history).
OpenCode: reads ~/.local/share/opencode/opencode.db directly.
Cursor: reads ~/Library/Application Support/Cursor/User/globalStorage/{conversation-search,state}.db. macOS only. Status is always unknown — no running signal is available.
Codex: reads ~/.codex/state_5.sqlite, falling back to scanning ~/.codex/sessions/**/*.jsonl directly if the DB schema doesn't match.
All formats are undocumented by their vendors. Adapters degrade (fewer fields) rather than crash on an unrecognized record.
setup subcommand and Cursor auto-registration are not implemented — register manually as shown above.
The plugin configs pin harness-homies@<version>. When releasing, bump the version in package.json, plugin/plugin.json, plugin/.claude-plugin/plugin.json, both MCP configs in plugin/, opencode/opencode.json, and src/server.ts.
Get a single session's title, cwd, status, todos, and message count. Use the agent and id fields from a list_sessions/search_sessions result. Read-only.
Inputs: agent (required), session_id (required).
Fetch a session's transcript, paginated (most recent messages by default). Secrets are redacted unless raw is set. Use the agent and id fields from a list_sessions/search_sessions result. Read-only.
Inputs: agent (required), session_id (required), raw, role, +2 more.
List coding-agent sessions on this machine across Claude Code, OpenCode, Cursor, and Codex — what's running, its topic, and its status. Read-only.
Inputs: cwd, agent, limit, running_only.
Search session titles (and recent transcript content) for a substring across one or all agents. Read-only, case-insensitive.
Inputs: query (required), agent, limit.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0get_sessionget_transcriptlist_sessionssearch_sessionsScored across 4 tools
Each tool targets a distinct concern: listing all sessions, retrieving one session's metadata, fetching a transcript, and searching across sessions. The overlap between list_sessions and search_sessions is minimal and clearly differentiated by behavior.
All tool names follow a consistent verb_noun snake_case pattern: list_sessions, get_session, get_transcript, search_sessions. The naming is predictable and instantly readable.
Four tools is well-scoped for a read-only session inspection server. Each tool covers a distinct aspect of the domain without redundancy or missing essentials.
For a read-only monitoring tool, the surface is complete: discover sessions, inspect one session, read its transcript, and search across sessions. No obvious lifecycle operations are expected here, so there are no meaningful gaps.
Experiment tracking for coding agents: read your team's ML runs, metrics and lineage. Read-only.
See which of your team's coding agents needs someone, across Claude Code, Codex and Cursor.
Read your epoc meeting transcripts: list synced meetings and read their transcripts. Read-only.
Read-only access to your CodeMouse accounts, repositories, and AI pull-request reviews.