agent-peek

Read-only visibility into your other AI agent sessions.

agent-peek lets one local agent ask what another local agent is doing without writing into its transcript, stealing focus, or rereading the whole session on every poll. It ships as a CLI, MCP server, and TypeScript library.

Why

When you run multiple coding agents in parallel, each one normally works in its own bubble. That leads to duplicated research, missed discoveries, and agents editing around each other without context.

agent-peek gives them a shared read-only window:

• Humans can browse active sessions with peek ui.
• Agents can call peek at <session> --json or the MCP tools.
• Scripts can poll cheaply with cursors via --since.
• Session transcripts are never modified.

Install

npm i -g agent-peek

Installed commands:

• peek — CLI
• apeek — alias for peek, useful if another tool owns that name
• agent-peek-mcp — MCP server for Claude Code, Codex, and other MCP clients

Quick Start

List discovered sessions:

peek list

Open the interactive browser:

peek ui

Peek at a session by display name, id, tag, or cwd:

peek at researcher-codex --mode structured

Get a compact local summary without an API key:

peek at researcher-codex --mode brief

Give a session a stable name:

peek tag researcher-codex as researcher
peek at researcher --mode summary

CLI

Common commands:

peek list                                 # show discovered sessions
peek list --adapter claude-code           # scan/list one adapter
peek list --all                           # include ended sessions
peek list --terminals                     # include tmux/screen terminal captures
peek list --ids                           # show raw session ids
peek list --json                          # machine-readable list
peek list adapters                        # show installed adapters
peek doctor                               # adapter availability and setup hints

peek ui                                   # interactive terminal browser
peek ui --adapter codex
peek ui --all
peek ui --terminals

peek at <name|id|tag|cwd>                 # raw snapshot
peek at <selector> --mode structured      # normalized status/task/tool fields
peek at <selector> --mode brief           # compact local summary, no API key
peek at <selector> --mode summary         # sentence-style summary, local by default
peek at <selector> --since <cursor>       # only messages since prior peek
peek at <selector> --first 20             # first 20 raw messages
peek at <selector> --last 50              # last 50 raw messages
peek at <selector> --around 100 --limit 30 # raw window around message 100
peek at <selector> --last 50 --reverse    # newest-first raw output
peek at <selector> --tools                # include tool-only raw messages

peek tag <selector> as researcher
peek untag researcher
peek register <adapter:id> at <path> [--as <name>]
peek forget <id>

Default peek list output is compact and human-first:

NAME               ADAPTER  STATUS  UPDATED  SOURCE  CWD
sessionseek-codex  codex    active  0s ago   file    ~/Documents/sessionseek/sessionseek

The NAME column is the selector to use with peek at. Raw ids stay available with peek list --ids, and JSON output includes both id and displayName.

Peek Modes

peek at supports four scriptable output modes:

| Mode | Use it for | API key | | --- | --- | --- | | raw | Reading transcript messages directly. Best for debugging or inspecting exactly what happened. | No | | structured | Stable fields for agents: current task, activity, last messages, pending tools, recent tools. | No | | brief | A compact local summary built from structured fields. Good default for humans and scripts that do not need raw logs. | No | | summary | Sentence-style summary. Uses a no-dependency local summary by default; can use Anthropic when explicitly configured. | No |

Raw mode has pagination controls:

peek at researcher --first 25
peek at researcher --last 100
peek at researcher --last 100 --offset 100
peek at researcher --around 250 --limit 40
peek at researcher --last 50 --reverse

By default, raw mode hides tool-only messages and tool-call status lines to keep the output readable. Add --tools or --verbose when you need that detail.

summary does not require Ollama, Anthropic, or any other model runtime. If you want hosted LLM summaries, set ANTHROPIC_API_KEY. To force local summaries on a machine that also has an Anthropic key, set:

AGENT_PEEK_SUMMARY_PROVIDER=local

Timeline is not a peek at --mode value. It is an interactive-only view inside peek ui.

Terminal UI

peek ui is for humans browsing in a real terminal. It shows a session list and a detail pane for the selected session.

It starts in structured mode. Press m or Tab to cycle through:

• structured — current task, activity, last messages, pending tools, recent tools
• brief — compact local summary, no API key
• timeline — chronological role/text timeline for quick scanning
• raw — recent transcript messages
• summary — sentence-style summary, local by default

There is no separate command-line flag for timeline yet; open peek ui, then press m/Tab until the header shows mode=timeline.

The detail pane shows useful metadata: raw id, adapter, source type, status, tag, cwd, transcript path, and last update time. It intentionally does not show cursors; cursors are for JSON/API callers that need incremental polling.

Keyboard controls:

• up/down or j/k — select a session
• Enter or Space — refresh the selected session detail
• m or Tab — switch detail mode
• r — rescan sessions
• q or Escape — exit

For pipes, scripts, and agent harnesses, use peek list, peek at, and peek at --json instead of peek ui.

Agent-Friendly Output

CLI failures are printed to stderr in a stable shape:

error: session_not_found
message: No session matched selector: worker
hint: Use `peek list` to get the current displayName values.
next:
  - peek list
  - peek list --ids
exitCode: 2

Cursor polling lets an agent fetch only new messages after a prior peek:

peek at researcher --mode raw --json
peek at researcher --mode raw --since <nextCursor> --json

MCP

The MCP server command is:

agent-peek-mcp

It is a local stdio server. Different clients use different config shapes.

Claude Code

Add it from the CLI:

claude mcp add agent-peek agent-peek-mcp

Or add a project-scoped .mcp.json:

{
  "mcpServers": {
    "agent-peek": {
      "command": "agent-peek-mcp"
    }
  }
}

Codex

Add to ~/.codex/config.toml:

[mcp_servers.agent-peek]
command = "agent-peek-mcp"

Cursor

Add to global ~/.cursor/mcp.json or project .cursor/mcp.json:

{
  "mcpServers": {
    "agent-peek": {
      "command": "agent-peek-mcp"
    }
  }
}

Windsurf

Add to ~/.codeium/mcp_config.json:

{
  "mcpServers": {
    "agent-peek": {
      "command": "agent-peek-mcp"
    }
  }
}

Gemini CLI

Add to user ~/.gemini/settings.json or project .gemini/settings.json:

{
  "mcpServers": {
    "agent-peek": {
      "command": "agent-peek-mcp"
    }
  }
}

Or use Gemini's MCP command:

gemini mcp add agent-peek agent-peek-mcp

Cline

Cline CLI uses ~/.cline/data/settings/cline_mcp_settings.json. The VS Code extension opens its own cline_mcp_settings.json from the MCP Servers settings.

{
  "mcpServers": {
    "agent-peek": {
      "command": "agent-peek-mcp",
      "disabled": false
    }
  }
}

VS Code

VS Code uses top-level servers, not mcpServers. Add to workspace .vscode/mcp.json or your user-profile mcp.json:

{
  "servers": {
    "agent-peek": {
      "type": "stdio",
      "command": "agent-peek-mcp"
    }
  }
}

Tools exposed:

• list_sessions
• peek_session
• tag_session

Config references: Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Cline, and VS Code.

Agent Skill

This repo includes an installable skill at skills/agent-peek. Use it when you want another agent to learn the peek workflow and the MCP configs above.

Install it with the npx skills CLI:

npx skills add akhileshrangani4/agent-peek

The installer is interactive and will guide scope/agent choices. For a non-interactive global install:

npx skills add akhileshrangani4/agent-peek --skill agent-peek -g -y

To target specific agents non-interactively:

npx skills add akhileshrangani4/agent-peek --skill agent-peek -a codex -a claude-code -g -y

Flags: -g installs globally for your user, and -y skips confirmation prompts. Omit -g if you only want the skill installed into the current project.

The skill teaches agents to:

• install or verify agent-peek
• run peek doctor, peek list, and bounded peek at commands
• configure MCP for Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Cline, and VS Code
• report what another session is doing without modifying that session

Library

import { createEngine } from "agent-peek";

const engine = await createEngine();
const sessions = await engine.list();

const result = await engine.peek("researcher", { mode: "summary" });
const firstPage = await engine.peek("researcher", { mode: "raw", from: "start", limit: 50 });

console.log(result.snapshot);
console.log(result.nextCursor);

Built-In Adapters

• claude-code — reads ~/.claude/projects/*/<uuid>.jsonl
• codex — reads Codex CLI transcripts in ~/.codex/sessions/<YYYY>/<MM>/<DD>/rollout-*.jsonl
• copilot-cli — reads GitHub Copilot CLI session-state directories in ~/.copilot/session-state/*
• gemini — reads Gemini CLI transcripts in ~/.gemini/tmp/<project>/chats/session-*.json
• goose — reads Goose session records in ~/.local/share/goose/sessions/sessions.db
• opencode — reads OpenCode filesystem storage in ~/.local/share/opencode/storage/{session,message,part}
• screen — captures GNU screen scrollback via hardcopy -h
• tmux — captures tmux pane output

Terminal adapters are opt-in for default CLI/MCP listing because they capture terminal scrollback, not structured agent transcript files. Use peek list --terminals or peek list --adapter tmux.

External Adapters

Set AGENT_PEEK_ADAPTER_PATH to a colon-separated list of adapter modules. Each module's default export must implement the Adapter interface from agent-peek/adapter.

Security Model

agent-peek is same-user, same-machine only.

• No remote transport is exposed by the package.
• Access control is your local filesystem permissions.
• Session access is read-only.
• agent-peek never writes to another agent's transcript.

License

MIT