Sidekick Agent Hub
See what your AI coding agents are doing and what they cost: live activity maps, tokens, cost, context and quota for Claude Code, Codex and OpenCode, in VS Code and your terminal.
- Source repo
- cesarandreslopez/sidekick-agent-hub
- Stars
- ★ 91
- Last updated
- 2d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 73/100 · Some gaps
At a glance
- Works with
- Universal · cross-platformCodex · Claude Code
- You'll need
- Typical use
- Heavy Claude Code users who want to know how many tokens each session burns and when context nears its limit
- Main limitation
- Session monitoring only supports Claude Code, Codex and OpenCode — other agents provide no data to read
- Website
- cesarandreslopez.github.io ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-19
- Source review
- 73/100 · Some gaps
What does this agent do, and when should you use it?
Sidekick Agent Hub (repository cesarandreslopez/sidekick-agent-hub, MIT-licensed, latest release 0.33.5) is an observability tool for AI coding agents, shipped as both a VS Code extension and a terminal CLI. It reads the session files that Claude Code, Codex and OpenCode already write — no proxy and no wrapper, so your workflow stays unchanged. Core capabilities include an eleven-station activity map (files, shell, network, memory, browser, subagents and more), token and cost tracking, context fill, five-hour billing blocks and plan quota projections, plus tasks, decisions and knowledge notes that carry across sessions. It also offers an account switcher for Claude Code and Codex logins without /logout cycles, a read-only MCP server your agent can query (sidekick mcp), a Claude Code status line (sidekick statusline), and a public state.json snapshot for tmux bars and scripts. The VS Code extension additionally runs AI features through Claude Max, the Claude API, OpenCode or Codex: inline completions, code transforms, commit messages, PR descriptions, code review and inline chat.
Sidekick parses local agent session files and produces live views and reports. The activity map routes every tool call to one of eleven stations (files, shell, network, memory, browser, agents, skills, hooks, plan, model, you), showing faults, subagents and the route the agent took; Insights show where time, errors and context went and catch retry streaks and edit ↔ test loops. Session analytics display tokens and cost per session and model, context fill, five-hour and weekly quota projections, the open billing block with its burn rate, and usage reports by day, week, month or session. The CLI provides sidekick dashboard (a full-screen view with nine panels), sidekick map --follow (live activity map), sidekick quota --all, sidekick daily, sidekick blocks, sidekick report (self-contained HTML report), sidekick dump --prompts, sidekick search and sidekick doctor. sidekick mcp starts a read-only MCP server so the running agent can check its own quota, burn rate, context pressure, activity map and project notes; sidekick statusline is a one-line footer for Claude Code that also records the official rate limits. sidekick accounts manages and switches logged-in Claude Code and Codex accounts, with VS Code auto-switching via sidekick.accounts.autoSwitchThreshold when quota crosses a threshold.
- Heavy Claude Code users who want to know how many tokens each session burns and when context nears its limit
- Developers on Pro/Max subscription quotas who need projections for the 5-hour billing block and weekly quota
- People working across multiple Claude Code / Codex accounts who want to switch without /logout cycles
- Developers who want tasks, decisions and handoff notes to survive into the next session instead of starting from zero
- Terminal-first engineers who want sidekick map --follow in a tmux split to follow agent activity live
- Users who want their agent to query its own quota and context pressure via MCP so it doesn't hit walls mid-task
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npm install -g sidekick-agent-hub
- Config files
~/.claude/settings.json- Runtime
- Node.js 20+
How do you install or deploy this agent?
VS Code: install the Sidekick Agent Hub extension from the VS Code Marketplace (ID: CesarAndresLopez.sidekick-for-max) or Open VSX, then open the Agent Hub icon in the activity bar.
Terminal requires Node.js 20+ (22.15+ to read compressed Codex sessions):
npm install -g sidekick-agent-hubThe npm package is sidekick-agent-hub; the command it installs is sidekick. No proxy or wrapper is needed — keep running claude, codex or opencode as usual; DB-backed OpenCode session monitoring expects an executable sqlite3 in the host environment.
How do you use this agent?
Launch the full-screen terminal dashboard (nine panels: Sessions, Tasks, Kanban, Notes, Decisions, Plans, Events, Charts, Map):
sidekick dashboardCommon one-shot commands:
sidekick map --follow
sidekick quota --all
sidekick daily
sidekick blocks
sidekick report
sidekick dump --prompts
sidekick search "migration"
sidekick doctor
sidekick accountsRegister the MCP server so the agent can query its own state:
claude mcp add sidekick -- sidekick mcp
codex mcp add sidekick -- sidekick mcpClaude Code status line — add to ~/.claude/settings.json:
"statusLine": { "type": "command", "command": "sidekick statusline" }In VS Code, use the Agent Hub sidebar: Session Analytics, Activity Map (open in an editor tab to scrub and replay), Accounts, Mind Map, Kanban Board, Plans, Project Timeline, Knowledge Notes, Subagents and Event Stream.
What are this agent's strengths and limitations?
- No proxy or wrapper: reads session files your agent already writes, keeping your claude/codex/opencode workflow unchanged
- One data source feeds a VS Code extension, terminal TUI, one-shot CLI commands, an MCP server and a public state.json snapshot
- The activity map routes tool calls across eleven stations and replays the agent's path; Insights catch retry streaks and edit ↔ test loops
- Quota projections cover the 5-hour billing block, burn rate and weekly quota, with account switching free of /logout cycles
- Session monitoring only supports Claude Code, Codex and OpenCode — other agents provide no data to read
- DB-backed OpenCode session monitoring depends on a sqlite3 executable being present on the host
- Costs nothing could price read as — rather than $0, so cost data depends on pricing information in session logs
- The MCP server is read-only: the agent can query its state but cannot control anything through it
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Sidekick Agent Hub This agent | 73 · Some gaps | — | ★ 91 | 2d ago | TypeScript | Codex · Claude Code |
| ccusage | 66 · Some gaps | — | ★ 20k | today | Rust | Codex · Claude Code |
| Ordewell | 87 · Good | CLIFree + model costs | ★ 211 | today | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| NeuroSquad Terminal Dashboard | 80 · Good | CLIFree + model costs | ★ 31 | 1mo ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Monitoring is read-only, the MCP server is explicitly read-only, no proxy/wrapper needed — least privilege is broadly satisfied; deducted because the account-switcher registers/switches Claude Code and Codex logins, touching credential state with no visible handling detail. Auto-switch (autoSwitchThreshold) changes accounts without explicit confirmation, so user confirmation is thin. Data flow: SECURITY.md claims no telemetry and local-only code, but its scope covers only the extension and its version table is stuck at 0.2.x. Dependency security: no lockfile/manifest evidence in the packet, only advice to keep dependencies updated — scored 1. External effects (state.json writes, login switching) are documented; rollback is explicit ('switches verified and can be undone'). Source attribution is strong: trawl contributor credited, THIRD_PARTY_NOTICES and CONTRIBUTORS present.
AGENTS.md, README and CI workflows are mutually consistent; doctor diagnostics, structured onDiagnostic reporting, and diagnostic-not-throw for missing sqlite3 show deliberate failure handling. Deduction: SECURITY.md supported-versions table (0.2.x) contradicts the shipped 0.33.5, hurting self-consistency. Dependency availability depends on Node 20/22 and an external sqlite3 runtime, both documented.
Audience and scenarios are clear (VS Code, terminal, scripted one-shot commands, multi-provider table). Environment fit is strong: Cursor conflict guidance, Node version notes, SEA build, Open VSX. Capability boundaries are partially documented (OpenCode sqlite3 note, provider matrix) but no unified limitations list. Trigger precision: on-demand completions with configurable debounce are well specified.
Information architecture is excellent: monorepo layout, docs site, architecture page, CLI guide links. Install notes cover Marketplace/Open VSX/VSIX/source/npm with verification. Naming stability loses a point: the Marketplace ID remains 'sidekick-for-max', a legacy artifact; npm-vs-command name difference is at least explained. Examples exist (walkthrough, CLI commands) but no FAQ is visible. Known limitations are scattered. MIT license is complete with third-party notices. Version 0.33.5 and changelog pointer are clear, but SECURITY.md versioning is stale. Maintenance is single-maintainer with stated response times that remain assertions.
Output usability is strong: activity map, quota projections, HTML report, scriptable state.json, honest 'never $0' pricing display. Marginal value is clear — zero-intrusion monitoring of files agents already write, plus context carry-over. Cost-benefit scored 2: claims of no proxy overhead cannot be verified statically.
Claims are traceable: CI badges, architecture docs, provider docs, a demo-recording script. Cross-source corroboration is good (README/AGENTS.md/CLAUDE.md/CI agree) minus the SECURITY.md contradiction. Fact/inference separation is deliberate: '—, never $0' for unpriced usage, and explicit provenance fields (PricingProvenance, TranscriptSourceProvenance).
- Account auto-switch (autoSwitchThreshold) changes saved logins automatically when quota crosses a threshold — disable or confirm before relying on it.
- SECURITY.md's supported-versions table (0.2.x) is stale versus 0.33.5, and the security policy scope covers only the extension package, not the CLI or shared library.
- Account switching touches Claude Code/Codex credential files; static review saw no handling detail — review the relevant source before use.
- OpenCode DB monitoring requires an executable sqlite3 runtime in the host environment; functionality degrades without it.
- The publisher is unverified and this is a single-maintainer project; dependency updates and security response carry no third-party assurance.
FAQ
Is Sidekick Agent Hub open source?
Is Sidekick Agent Hub safe to use?
Is Sidekick Agent Hub still maintained?
Who makes Sidekick Agent Hub?
What is Sidekick Agent Hub's official website?
Thanks — we use this to improve the page.