lm-assist

The Observability Platform for Claude Code & Agent SDK

Monitor, debug, and control AI coding agents with full session visibility, real-time execution tracking, and 155+ REST API endpoints.

• Monitor — real-time execution tracking, per-model cost & token breakdown, SSE event stream
• Debug — 15 insight views per session: Chat, Thinking, Agents, Plans, Team, DAG, Files, Git & more
• Control — web terminal from any browser, start/abort agents via API, remote access from anywhere

Install

/plugin marketplace add langmartai/lm-assist
/plugin install lm-assist@langmartai

Then open a new Claude Code session and run /assist-setup.

Read: Inside Claude Code: The Session File Format and How to Inspect It — technical breakdown of the JSONL session format, message types, subagent trees, and how lm-assist surfaces it all.

Why lm-assist

Claude Code and the Agent SDK have no built-in dashboard. You get a terminal or logs. When you're running multiple agents, debugging a failed execution, or tracking costs across a fleet of machines — you need full visibility into every session, every subagent, every tool call, every token spent.

| Without lm-assist | With lm-assist | |-------------------|---------------| | Scroll through terminal output | 15 specialized views per session | | No cost visibility | Per-model token & cost breakdown | | Can't see what agents are doing | Real-time execution dashboard | | No way to inspect subagent trees | Full DAG visualization | | Terminal-only access | Web UI from any device, anywhere | | Agent SDK runs are black boxes | Same session inspection as CLI |

Three Pillars

1. Monitor

Real-time execution tracking via REST API.

• Session list with slug names, live status, running process detection
• Per-session and per-project cost tracking with per-model token breakdown
• SSE event stream for real-time updates (GET /stream)
• Multi-machine fleet dashboard via LangMart Hub
• Statusline: context %, rate limits (5h/7d), cost, RAM, PID

Key endpoints: GET /monitor/executions · GET /stream · GET /sessions · GET /projects/sessions

2. Debug

15 data dimensions per session, all accessible via API.

• Conversation, thinking blocks, tool calls, subagent hierarchy, DAG
• File changes, git operations, plans, tasks, team coordination
• Fork tracking, session summaries, skill traces

Key endpoints: GET /sessions/:id · GET /sessions/:id/dag · GET /sessions/:id/subagents · GET /sessions/:id/conversation

3. Control

Full runtime management API.

• Start and abort agent executions
• SDK runner for programmatic headless execution
• Session cache warm/clear
• Web terminal (ttyd) from any browser
• Remote access via LangMart Hub

Key endpoints: POST /monitor/abort/:executionId · POST /ttyd/start · POST /agent/execute · POST /hub/connect

Web Dashboard

See claude-code-webui for the full web dashboard with 15 insight tabs, web terminal, task kanban, and mobile support.

Data Sources

lm-assist reads the same JSONL session files regardless of how they were created:

| Source | What It Produces | lm-assist Coverage | |--------|-----------------|-------------------| | Claude Code CLI | Interactive sessions with subagents, teams, worktrees | Full parsing — all message types, tool calls, metadata | | Agent SDK (Python/TypeScript) | Programmatic agent executions | Same JSONL format — full session inspection | | Headless mode (claude -p) | Background/CI runs | Detected via process status store | | Running processes | Live PID, tmux, terminal state | Real-time monitoring with zero polling overhead |

Install

Quick start

npm install -g lm-assist

The postinstall script automatically starts services, installs the statusline, and installs the Claude Code Multisession plugin.

Open a new Claude Code session and type /sessions to verify.

Plugin marketplace

Add the marketplace once — then install any combination of plugins:

/plugin marketplace add langmartai/lm-assist

Three packages work together:

┌─────────────────────────────────────────────────────────┐
│  claude-code-multisession          (Claude Code plugin) │
│  Skills: observe, route                                 │
│  Commands: /projects /sessions /summary /run            │
├─────────────────────────────────────────────────────────┤
│  claude-code-webui                 (Claude Code plugin) │
│  Skill: dashboard                                       │
│  Commands: /web /web-sessions /web-tasks                │
├─────────────────────────────────────────────────────────┤
│  lm-assist                              (npm package)   │
│  Foundation: 155+ API endpoints, Next.js web dashboard, │
│  session engine, cost tracking, statusline              │
│  Commands: /assist /assist-setup /assist-status ...     │
└─────────────────────────────────────────────────────────┘

lm-assist is the foundation — the API server and web dashboard that powers everything. The plugins add skills and commands on top.

| Install command | Layer | What you get | |----------------|-------|-------------| | npm install -g lm-assist | Foundation | API server (:3100), web dashboard (:3848), statusline, 155+ endpoints | | /plugin install claude-code-multisession@langmartai | Skills | observe + route skills, /projects /sessions /summary /run | | /plugin install claude-code-webui@langmartai | Web access | dashboard skill, /web /web-sessions /web-tasks | | /plugin install lm-assist@langmartai | Setup & diagnostics | /assist-setup /assist-status /assist-search /assist-logs |

Install all three for the full experience, or pick what you need.

Install from source

git clone https://github.com/langmartai/lm-assist.git
cd lm-assist
npm install && npm run build
./core.sh start

Then in Claude Code: /plugin install .

Services

| Service | Port | Description | |---------|------|-------------| | Core API | 3100 | REST API — sessions, monitor, agents, tasks, knowledge | | Web UI | 3848 | Next.js dashboard — 15 insight tabs, terminal, settings |

lm-assist start       # Start both services
lm-assist stop        # Stop all services
lm-assist status      # Health check + process info
lm-assist upgrade     # Upgrade to latest version

Skills & Commands

Skills and commands are provided by the Claude Code Multisession plugin — installed automatically via /assist-setup.

Skills (auto-triggered via Claude Code Multisession)

Skills activate automatically when Claude detects relevant intent — no slash command needed.

| Skill | Triggers on | What it does | |-------|------------|--------------| | observe | "what's running?", "session costs", "show subagents", "run this on project X" | Full observability — monitor sessions, debug agents, control executions, manage summaries | | route | Prompt mentions another project's features or codebase | Cross-project routing — evaluates whether to stay, resume, fork, queue, or start new |

Commands

| Command | Description | |---------|-------------| | /sessions | Session list with costs, turns, running status | | /summary | Summarize current session, generate display name, record learning | | /run <prompt> | Execute an agent with pre-flight checks | | /assist | Open the web UI in your browser | | /assist-status | Show status of all components | | /assist-setup | Start services and verify integrations | | /assist-search | Search the knowledge base | | /assist-logs | View context-inject hook logs | | /assist-mcp-logs | View MCP tool call logs |

Use Cases with Examples

"What sessions are running and how much have they cost?"

> /sessions
Sessions (3 running, 415 total)
───────────────────────────────────────────────────────────────────────
Status  Name                         Project            Model      Cost  Turns
───────────────────────────────────────────────────────────────────────
[RUN]   observability-platform-build  lm-assist          opus    $307.66  1462
[RUN]   trade-delta-analysis          my-trading-app   opus    $153.20   822
[RUN]   anti-kelly-system             my-trading-app   opus     $35.49   433
        two-track-dashboard           my-trading-app   opus    $107.43   516
        skill-validation-test         lm-assist          opus      $0.63    11
───────────────────────────────────────────────────────────────────────
Total cost: $604.41

"Summarize what this session has been doing"

> /summary
Session Summary
═══════════════
Name:    observability-platform-build
Project: lm-assist
Turns:   1462 | Cost: $307.66
Status:  in progress

What this session is about:
  Building the lm-assist observability platform — session summaries,
  cross-project routing, prompt queue, auto-learning.

What was accomplished:
  • Session slug/customTitle parsing from Claude Code JSONL
  • Knowledge system kill switch (saves ~100MB)
  • Observability skill + routing skill + 3 new commands
  • Persistent session summaries with delta tracking
  • Universal prompt queue with source/target context
  • Auto-learning signals that evolve routing over time

"Fix the delta analysis bug in the trading pipeline"

The route skill auto-detects this belongs to another project:

This task appears to belong to my-trading-app, not the current project.

Relevant session found: trade-delta-analysis
Summary: Implemented delta analysis mode for /analyze command...
Status: idle (not running)

Recommendation: RESUME
Reason: This session built all existing delta modes and knows the spec,
flag table, and pre-launch copy patterns.

To resume:  claude --resume def67890-session-id

"Run a security review on this project"

> /run review all API endpoints for security vulnerabilities
Checking running executions... OK: none running.

Execution started: agent-1774468578687-1p2cnh
  Session: nifty-wishing-boole
  Status: running
  Monitor with: /sessions
  Web UI: http://localhost:3848

"What's the cost breakdown by project?"

The observe skill activates:

Project costs:
  lm-assist              $308.29  (80 sessions)
  my-trading-app       $892.15  (440 sessions)
  my-web-platform          $45.20  (12 sessions)
  my-agent-framework              $23.50  (15 sessions)

Session Intelligence

lm-assist learns from every interaction to route prompts smarter over time:

• Session summaries — LLM-generated descriptions stored with delta tracking. Updated incrementally as sessions grow.
• Project summaries — Comprehensive reference generated by agents that explore each project's CLAUDE.md, scripts, configs, and codebase.
• Prompt queue — When a session is busy, new work is queued with full context (original intent, formatted prompt, routing reason, context hint).
• Auto-learning — Keywords, commands, and routing patterns accumulate with frequency counts. After a few interactions, routing skips deep scans entirely.

Learning context for my-trading-app:
  Frequently mentioned: delta analysis(3x), anti-kelly(2x), regime analysis(1x)
  Areas worked in: analysis pipeline(1x)
  Routing patterns: delta analysis → trade-delta-analysis session

Multi-Project Concurrent Development

lm-assist shines when you're working across multiple projects simultaneously. The routing skill, session summaries, and prompt queue work together to keep context organized.

Scenario: You're building lm-assist while also maintaining a trading platform and a web gateway.

1. Start your day — see everything at a glance:

> /sessions
Sessions (5 running, 547 total)
───────────────────────────────────────────────────────────────────
Status  Name                         Project            Model      Cost  Turns
───────────────────────────────────────────────────────────────────
[RUN]   observability-platform-build  lm-assist          opus    $355.84  1573
[RUN]   trade-delta-analysis          my-trading-app   opus    $153.20   843
[RUN]   anti-kelly-system             my-trading-app   opus     $35.49   433
        two-track-dashboard           my-trading-app   opus    $107.43   516
        gateway-ssl-upgrade           my-web-platform     opus     $12.30    85
───────────────────────────────────────────────────────────────────
Total cost: $664.26

2. You're in lm-assist and say "fix the regime analysis stale data bug"

The route skill detects this belongs to another project:

This task belongs to my-trading-app (regime analysis is part of the trading pipeline).

Relevant session: trade-delta-analysis
Summary: Implemented delta analysis mode for /analyze. Knows all delta
patterns, flag tables, and pre-launch copy rules.
Status: running (T:843, $153.20)

Recommendation: QUEUE (session is running)
Queued as high priority. Will process when current work completes.

Queue for trade-delta-analysis (2 pending):
  [high]   fix the regime analysis stale data bug
  [normal] add S/R interaction delta mode

3. Back in lm-assist, you say "add error handling to the learning store"

The route skill checks the current session summary:

STAY — "learning store" matches current session work (observability-platform-build).
The session summary mentions auto-learning, learning signals, and learning store.

No routing. Work continues in the current session.

4. You say "the gateway SSL cert is expiring, renew it on SG"

The route skill identifies this as a simple remote operation:

This involves my-web-platform (gateway) on the SG server.
But it's a simple operations task — can handle from here via SSH.

Running: ssh user@prod-server 'sudo certbot renew'

No project switch needed.

5. You say "deploy the trading dashboard changes to production"

Route skill identifies the target:

This belongs to my-trading-app (trading dashboard).

Relevant session: two-track-dashboard (completed, 516 turns)
Summary: Built two-track cross-product trading dashboard with spread
and cross-product opportunity tables.

Recommendation: RESUME
Reason: This session built the dashboard and knows the deployment process.

To resume: claude --resume abc12345-session-id

6. End of day — summarize what happened:

> /summary
Session Summary: observability-platform-build
  Today: added error handling to learning store, fixed 3 bugs,
  queued 2 tasks for trade-delta-analysis, deployed gateway cert.

> /sessions my-trading-app
  trade-delta-analysis: processed 2 queued prompts (regime fix, S/R delta)
  two-track-dashboard: deployed to production
  anti-kelly-system: idle since yesterday

Key patterns for multi-project work:

| Situation | What happens | |-----------|-------------| | Task for current project | Route skill says STAY — no context switch | | Task for another project, relevant session running | QUEUE — prompt waits, processes when session is free | | Task for another project, session idle | RESUME — switch to that session with full context | | Simple cross-project operation (ssh, curl, read) | Handle locally — no project switch | | Task for unknown project | Route skill asks for clarification | | Same task mentioned again | Learning signals shortcut the routing — instant match |

How learning makes it faster over time:

First time: "fix delta analysis" → scan all project summaries → find my-trading-app → 4 API calls

After learning: "fix delta analysis" → signal says delta analysis(5x) → my-trading-app → 1 API call

The more you work across projects, the smarter routing gets. Keywords, commands, and routing patterns accumulate automatically.

Key API Endpoints

| Category | Endpoints | Highlights | |----------|-----------|------------| | Sessions | 27 | List, detail, delta fetch, batch-check, conversation, subagents, forks, DAG | | Monitor | 6 | Running executions, summary, abort, SSE stream | | Projects | 12 | List projects, sessions per project, git info, worktree detection, costs | | Terminal | 13 | ttyd start/stop/status, WebSocket proxy, tmux attach | | Tasks | 22 | Task lists, aggregated tasks, ready tasks, dependency tracking | | Summaries | 10 | Session summaries, project summaries, needs-update check | | Queue | 10 | Prompt queue with source/target, priority, dispatch/complete lifecycle | | Learning | 4 | Record signals, query by project, learning context generation | | Search | 4 | Session content search, recent sessions, vector search | | Skills | 9 | Skills analytics, chains, traces, per-session breakdown | | Knowledge | 21 | List, search, generate, review (optional — can be disabled) |

All endpoints support ifModifiedSince for efficient polling. Session data supports three indexing dimensions: lineIndex (JSONL position), turnIndex (conversation turn), and userPromptIndex (user message count).

Configuration

No API key needed — lm-assist works entirely with your local Claude Code session data. Optionally copy .env.example to .env:

API_PORT=3100                    # Core API port (default: 3100)
WEB_PORT=3848                    # Web UI port (default: 3848)

Platform Support

| Platform | Support | Notes | |----------|---------|-------| | Linux | Full | All features including web terminal | | macOS | Full | All features including web terminal | | Windows | Partial | Everything except console/terminal access (ttyd not available) | | Mobile / Tablet | Web UI | Monitor, debug, and control from any device on your network |

Who It's For

• Solo developers using Claude Code — see what's happening across all your sessions
• Teams building with Agent SDK — observability for your agent pipelines
• DevOps managing agent fleets — multi-machine dashboard, cost tracking, process management
• AI product builders — debug agent behavior with 15 insight views

Changelog

See CHANGELOG.md for version history and release notes.

Requirements

• Node.js >= 18
• Claude Code (for slash commands and MCP integration)

Related

• claude-code-multisession — Skills plugin: cross-project session routing, /projects, /sessions, /summary, /run
• claude-code-webui — Web dashboard plugin: 15 insight tabs, web terminal, /web, /web-sessions, /web-tasks
• Knowledge system — Optional: auto-extract knowledge from sessions, MCP tools, context injection. Off by default, enable in Settings > Data Loading

License

AGPL-3.0-or-later