multi-agent-room

A minimal Claude Code conversation viewer — runs as an HTTP proxy in front of the Anthropic API, captures every /v1/messages request/response, and shows the conversation (user / assistant / tool_use / tool_result / thinking) in a local web UI.

Inspired by cc-viewer; this is a stripped-down, single-file-per-concern version.

How it works

 Claude Code ──HTTP──▶ multi-agent-room :7080 ──HTTPS──▶ api.anthropic.com
                         │
                         ├─ records request + assembled SSE response
                         │  to ~/.mar/logs/<sessionId>.jsonl
                         │
                         └─ pushes new entries via SSE to the React UI

Install

git clone https://github.com/weiesky/multi-agent-room.git
cd multi-agent-room
npm install              # runs `vite build` automatically via the prepare script

Usage

Terminal A — start the server:

# After `npm install -g .` (or `npm link`):
mar

# Or, from source without a global install:
node bin/multi-agent-room.js

The CLI is also installed as multi-agent-room for backward compatibility, but mar is the canonical short command going forward.

You'll see:

✓ mar running at http://127.0.0.1:7080

  In another terminal:
    export ANTHROPIC_BASE_URL=http://127.0.0.1:7080
    claude

  Then open http://127.0.0.1:7080 to view conversations.

Terminal B — point Claude Code at the proxy:

export ANTHROPIC_BASE_URL=http://127.0.0.1:7080
claude

Browser: open http://127.0.0.1:7080. The current session shows up in the left list with a live tag and updates in real time.

⚠ If you also have ANTHROPIC_BASE_URL set in ~/.claude/settings.json, that takes precedence over the shell export. Either remove it from settings.json, or edit it to point at this proxy.

Development

npm run dev   # starts the server (:7080) and Vite dev server (:5173) together

Vite proxies /api, /events, and /v1 to the backend, so opening http://127.0.0.1:5173 gives you HMR with live data.

Environment variables

| Variable | Default | Description | |---|---|---| | PORT | 7080 | Server port. The server scans 7080–8000 and binds the first free port. PORT, if set, must fall in that range and acts as the start of the scan; out-of-range values are warned about and ignored. The actual bound port is printed in the startup banner. | | MULTI_AGENT_ROOM_LOG_DIR | ~/.mar/logs | Where session JSONL files are written | | MULTI_AGENT_ROOM_UPSTREAM | https://api.anthropic.com | Upstream Anthropic /v1/messages API to forward to | | MULTI_AGENT_ROOM_OPENAI_UPSTREAM | (none) | Upstream host for the OpenAI Responses API used by Codex CLI. Falls back to OPENAI_UPSTREAM_URL → MULTI_AGENT_ROOM_UPSTREAM → https://api.openai.com. Required when codex's [model_providers.<name>].base_url points to a non-OpenAI host (e.g. a third-party provider like aicodewith / cch); otherwise the proxy will forward /chatgpt/v1/responses to api.openai.com and every request will fail. | | MULTI_AGENT_ROOM_AGENTS_FILE | ~/.mar/agents.json | Path to the agents config file |

Agents (stage 1)

The same server hosts a small agent configuration store at ~/.mar/agents.json (next to logs/). Open the Agents tab in the web UI to manage it, or use the JSON API:

# List
curl http://127.0.0.1:7080/api/agents

# Create
curl -X POST http://127.0.0.1:7080/api/agents \
  -H 'Content-Type: application/json' \
  -d '{"name":"my-claude","type":"claude-code","workspace":"/Users/me/proj","env":{"FOO":"bar"}}'

# Get / patch / delete
curl http://127.0.0.1:7080/api/agents/my-claude
curl -X PATCH http://127.0.0.1:7080/api/agents/my-claude \
  -H 'Content-Type: application/json' -d '{"workspace":"/new/path"}'
curl -X DELETE http://127.0.0.1:7080/api/agents/my-claude

Each agent record stores: name (unique, ^[A-Za-z0-9_-]{1,64}$), type (one of openclaw / claude-code / codex / hermes / gemini-cli), workspace (absolute path), env (string→string), and three free-form JSON fields: mcp, skills, ext_info. Stage 1 only stores; spawning agents and @name routing arrive in stage 2 — by convention, stage-2 spawn parameters live under ext_info.spawn = { command, args, cwd }.

The store atomically writes agents.json.tmp → rename and keeps a single-step agents.json.bak; corruption recovers from .bak, or archives the broken file as agents.json.broken-<ts>.

Don't run two multi-agent-room processes against the same ~/.mar/ — there's no inter-process lock. One server at a time.

What's stored

~/.mar/logs/<sessionKey>-<YYYYMMDD>-<seq>.jsonl — one JSON object per line, one line per /v1/messages request/response pair:

{
  "id": "uuid",
  "sessionId": "37a8eec1-20260428-001",
  "ts": "2026-04-28T15:30:00.000Z",
  "method": "POST",
  "url": "/v1/messages",
  "status": 200,
  "durationMs": 1234,
  "preflight": false,
  "request": {
    "headers": { /* x-api-key, authorization, cookies removed */ },
    "body": { "model": "...", "messages": [...], "tools": [...], "system": "..." }
  },
  "response": {
    "role": "assistant",
    "content": [ {"type":"text",...}, {"type":"tool_use",...} ],
    "usage": {...}
  }
}

sessionKey is derived from metadata.user_id if present, otherwise from a hash of the system prompt — so different cwds / Claude Code instances get different sessions automatically.

What's intentionally not here

cc-viewer's full feature set is much larger. This project deliberately omits:

• Markdown rendering, syntax highlighting, mermaid
• Terminal PTY, statistics charts, Electron, mobile layout
• Multi-account / proxy profile switching, hot-reload
• Hooks / plugin system
• Delta storage / incremental log compaction
• MainAgent / subagent classification

If you need any of those, use cc-viewer.

Releasing

A push of a v* tag triggers .github/workflows/release.yml:

1. Build dist/ and pack a tarball with npm pack
2. Create a GitHub Release with auto-generated notes (tarball attached)
3. Run npm publish --provenance if NPM_TOKEN is configured as a repo secret

Pre-flight checklist (do once, before the first release):

• [ ] Add NPM_TOKEN (an Automation token from https://www.npmjs.com/settings/~/tokens) under the repo's Settings → Secrets and variables → Actions. Without it the workflow still creates the GitHub Release but silently skips the npm step.
• [ ] Confirm the package name multi-agent-room is owned by your npm account (or available — first publish claims it).
• [ ] npm pack --dry-run locally to eyeball the tarball contents.

# bump version, tag, push
npm version patch  # or minor / major
git push --follow-tags

License

MIT