Turn your vault into an
AI command center

Agent Fleet is an Obsidian plugin for building and running autonomous AI agents — on Claude Code or OpenAI Codex. Schedule tasks, chat in real time, connect Slack, Telegram, and Discord. Every agent is markdown you own.

Install in Obsidian View on GitHub

Agent Fleet / Dashboard 12 today

Active agents 5/5 5 of 5 enabled

Runs today 12 12 passed · 0 failed

Tokens used 48.2k $0.94 today

Scheduled tasks 4 Next: jira-digest in 2h

Run activity 14d

Success rate

95%19/20 runs

Recent activity

Fleet status

Backlog 2

summarize-meetingstask-manager

refactor-weekly-notevault-librarian

Scheduled 3

jira-digestodp-product-manager · 9:00

slack-digestodp-product-manager · 17:00

vault-weekly-organizevault-librarian · Sun

Running 1

prioritize-taskstask-manager · 2m 14s

Done 3

uptime-checksite-monitor · 38s

fleet-healthfleet-orchestrator · 52s

slack-digestodp-product-manager · 1m 45s

Failed 0

Nothing failed today

Backends

Runs on the subscriptions
you already pay for

No separate API bill required. Each agent picks its engine — mix freely in the same fleet. Chat, tasks, heartbeat, channels, and memory work identically on both.

Claude Code

@anthropic-ai/claude-code

Default

Works with a Claude Max or Pro subscription — or an Anthropic API key. Aliases opus / sonnet / haiku, pinned IDs, Bedrock, Vertex, Foundry.

OpenAI Codex

@openai/codex

Works with a ChatGPT Plus/Pro plan — or an OpenAI API key. Set an agent's adapter to Codex and everything else stays the same: pick the engine, keep the workflow.

Features

A fleet, not a chatbot

Agents with their own prompts, permissions, schedules, and memory — working while you're not asking.

agents/site-monitor/
├── agent.md       identity & system prompt
├── config.md      model, timeout, permissions
├── SKILLS.md      shared skill library
└── HEARTBEAT.md   autonomous schedule

Agents as folders of markdown

Every agent is a folder in your vault: system prompt, model, permission mode, allow/deny lists, skills, memory. Edit the files or use the UI — same thing.

0 */6 * * *

Heartbeat

Define what an agent does when no one is asking — monitoring, reports, health checks — with results posted straight to a channel.

Summarize today's Slack activity

↳ inline thread · own session

Interactive chat

Dock a chat panel anywhere. Stream responses, steer mid-task, attach docs and images. Inline threads let you tangent without polluting the main conversation.

Backlog

Scheduled

Running

Done

Task board

Kanban with cron scheduling, priorities, real-time progress, and abort. Route a nightly summary to Haiku while the agent stays on Opus.

Slack Telegram Discord

Message your fleet from anywhere

Full agent sessions over Slack, Telegram, and Discord — @agent-name routing, typing indicators, image attachments, sessions that survive restarts.

working.mdcurated · token-budgeted · injected every run

↑ nightly reflection consolidates

raw/2026-07-01.mdappend-only ground truth

Memory that curates itself

Agents record facts via a remember tool. A nightly reflection dedups, resolves contradictions, and can even propose new skills from recurring patterns — approval-gated.

_fleet/mcp/

claude codex

MCP, registered once

One registry, every agent, both backends. One-click OAuth 2.1 with tokens in your OS keychain. Your native configs are never touched.

inbox/ → _topics/ ⇄ Q&A

Wiki Keeper

Turn any folder into a self-maintaining wiki. Drop sources in an inbox; a scoped keeper agent files them into an interlinked topic tree with citations. Answers compound back in.

95% · 18/19 runs

Full observability

Run charts, success rates, streaming output from active agents, and token & cost tracking across every run — tasks, heartbeats, chat, and channels.

22 built-in skills Per-task model override Permission modes & allow/deny lists Run history logs Session resume Catch-up on missed tasks OS-keychain secrets

Philosophy

If the plugin disappears,
your knowledge stays

Every agent, skill, task, run log, and memory is a plain markdown file in your vault. Searchable, diffable, version-controllable — fully yours.

No proprietary database. No lock-in.
Edit any file by hand; the UI reflects it.
Secrets live in your OS keychain, never in the vault.
Sync with git, Obsidian Sync, or anything else.

_fleet/
├── agents/        one folder per agent
├── skills/        22 built-in, add your own
├── tasks/         frontmatter + cron
├── channels/      Slack · Telegram · Discord
├── mcp/           one file per server
├── runs/          every execution, logged
└── memory/        working set + raw archive

Quick start

Running in three steps

Install the plugin

Settings → Community plugins
→ Browse → "Agent Fleet"

Right from Obsidian's community plugin directory. Prefer the terminal? npm install -g obsidian-agent-fleet works too.

Install a backend

npm install -g @anthropic-ai/claude-code

Claude Code (Max/Pro subscription or API key) — or OpenAI Codex with a ChatGPT plan.

Meet your orchestrator

"Create an agent that summarizes
 my Slack channels every morning"

Enable the plugin and chat with the built-in Fleet Orchestrator — it creates agents, tasks, and skills for you.

FAQ

Questions

Do I need an API key?

Not necessarily. Agent Fleet works with your Claude Max or Pro subscription via the Claude Code CLI — no separate billing. An Anthropic API key works too, as does a ChatGPT plan for Codex agents.

Can I mix Claude and Codex agents?

Yes. The adapter is set per agent, so some agents can run on Claude Code and others on Codex in the same fleet. Chat, tasks, heartbeat, channels, and memory work the same on both.

What happens if I delete the plugin?

Your _fleet/ folder stays. All agents, tasks, skills, run logs, and memory are plain markdown files in your vault.

Does the Slack / Telegram / Discord bot work when Obsidian is closed?

No — channels run inside Obsidian, so the bot goes offline when Obsidian is closed. All connections are outbound (Socket Mode, long-poll, WebSocket), so no public URL is needed.

Can multiple agents run at the same time?

Yes, up to a configurable concurrency limit (default 2). Additional tasks queue until a slot opens.

Is my data safe?

All secrets — MCP OAuth/bearer tokens and channel credentials — are stored in your OS keychain via Obsidian's SecretStorage API, never in the vault. Agents respect per-agent permission modes and allow/deny lists.