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