eleven

KO | EN

gitlite — search

#artificial-intelligence #chatgpt #ai-agent #android #cross-platform #react-native #database #csharp #reactjs #http

(★ 9)

pi from anywhere

File Explorer

ZIP Download

- dashboard-threads.jpg
- dashboard-workspaces.jpg
- extending-the-status-ui.md
- vendor.mjs
- - claude-code.ts
  - claude-session-state.ts
  - host-api.ts
  - pi.ts
  - runner.ts
  - system-prompt.ts
  - task-activity.ts
  - tool-policy.ts
- - server.ts
- - conversation.ts
  - gc.ts
  - pending.ts
  - reader.ts
  - request-log.ts
  - search.ts
  - store.ts
  - usage.ts
- cli-client.ts
- cli.ts
- config.ts
- daemon.ts
- doctor.ts
- gateway.ts
- log.ts
- media-store.ts
- paths.ts
- provider-usage.ts
- schedules.ts
- service.ts
- util.ts
- claude-code-provider.test.ts
- dashboard-dom.test.ts
- dashboard-http.test.ts
- dashboard-live-turn.test.ts
- dashboard-markdown.test.ts
- dashboard-media.test.ts
- dashboard-message-display.test.ts
- dashboard-nav-drag.test.ts
- dashboard-pwa.test.ts
- dashboard-thread-pane.test.ts
- failover.test.ts
- gateway-live-turn.test.ts
- host-api.test.ts
- operational-cli.test.ts
- prose-blocks.test.ts
- schedules.test.ts
- task-activity.test.ts
- telegram-bot.test.ts
- thread-search.test.ts
- thread-store.test.ts
- tool-policy.test.ts
- transcribe-media.test.ts
- usage-stats.test.ts
.gitignore
CLAUDE.md
LICENSE
package-lock.json
package.json
README.md
tsconfig.build.json
tsconfig.json

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

A featherweight personal AI gateway. Channels, model providers, workspaces, threads, and scheduled turns.

eleven connects chat channels to agents running in your local directories, with a web dashboard to configure everything and browse every conversation. It uses the [pi SDK](https://github.com/badlogic/pi-mono) for 30+ model providers and can also run the official Claude Code runtime through Anthropic's Agent SDK, reusing a local Claude subscription login. Born as a lighter alternative to [OpenClaw](https://github.com/openclaw/openclaw).

The idea is simple: eleven makes your local agent reachable from anywhere, through any chat channel you wire up, so you can talk to it from your phone or a group when you are away from your machine. It keeps conversation and schedule execution state, but the agent's instructions stay in the workspace: `AGENTS.md`, skills, and scheduled prompts reach both runtimes, while Pi-backed models also retain Pi extensions unchanged.

![eleven dashboard: live conversation view](https://raw.githubusercontent.com/ceifa/eleven/main/assets/dashboard-threads.jpg)

## What it does

- **Channels**: connect your agent to any channel, including Telegram and Discord.
- **Pairing**: deny-by-default everywhere. Unknown DM senders and unregistered groups show up as requests in the dashboard; one click adds them to the allowlist.
- **Workspaces**: a directory + tool policy (`read`-only to full `bash`) + optional model override. pi loads the workspace's `AGENTS.md` and `.agents/skills` natively, so your agent's persona lives in the repo, not in eleven.
- **Dashboard** (`127.0.0.1:1111`): configure channels, providers, workspaces, approve pairings, and browse or join any thread with live streaming. Threads are named after the conversation they belong to — the forum topic, the group, the person — searchable full-text across every transcript, stoppable mid-turn, and restartable in place — "New thread here" is `/new` for the conversation you are reading, so a forum topic can be started over without leaving the browser — and every thread can show the skills its workspace loads. Its composer attaches files (pick, drop, or paste a screenshot) and records voice messages, transcribed the same way a Telegram voice note is.
- **On a phone**: the same dashboard on a 390px screen — the list and the conversation become two screens the back gesture moves between, the composer stays above the keyboard, and it installs to the home screen as a PWA (Settings → App). Its interface is cached, so it opens with no signal; the conversations in it still need the daemon.
- **Models**: any provider pi supports, plus the official Claude Code runtime.
- **Scheduled turns**: point a workspace at Markdown manifests with five-field cron expressions. eleven reads the prompt from the file at run time, records each run as a thread, catches up after downtime, and serializes jobs that share a workspace.
- **Usage**: where the tokens actually went — per day, per model and per thread, with the cache hit rate, the list price at each provider's published rates, and how many prompt tokens came in cold because a turn idled past the cache TTL or fell down the sequence onto another model. Read off the transcripts on disk, so there is no meter to keep in sync, and whatever quota windows a provider reports sit on the same page.

![eleven dashboard: workspace configuration](https://raw.githubusercontent.com/ceifa/eleven/main/assets/dashboard-workspaces.jpg)

## Install

Requires Node ≥ 24. Pi-backed models use [pi](https://pi.dev) credentials (`pi` → `/login`, or API keys in `~/.pi/agent/auth.json`). Claude Code models reuse the official local login:

```sh
claude auth login
```

```sh
npm i -g @ceifa/eleven
eleven init
eleven doctor
```

Config lives in `~/.config/eleven/eleven.json`, but the dashboard edits it for you, so the fastest path is: `eleven start`, open `http://127.0.0.1:1111`, add a bot token and a workspace.

## Configuration

```json
{
  "dashboard": { "port": 1111, "host": "127.0.0.1" },
  "models": [
    { "model": "openai-codex/gpt-5.6-sol", "reasoning": "high" },
    { "model": "claude-code/opus", "reasoning": "high" }
  ],
  "transcription": { "command": "whisper --file {{file}}" },
  "session": { "idleDays": 7, "retentionDays": 30 },
  "workspaces": {
    "agent": {
      "path": "~/my/agent-repo",
      "tools": ["read", "bash", "edit", "write"],
      "schedules": [
        {
          "path": "**/.meta/crons/*.md",
          "timeZone": "America/Sao_Paulo",
          "modelAliases": { "claude-opus-5": "claude-code/opus" }
        }
      ],
      "channels": [
        {
          "type": "telegram",
          "name": "main",
          "token": "$TELEGRAM_BOT_TOKEN",
          "users": { "123456789": { "name": "me" } },
          "groupAllowedUsers": [123456789],
          "groups": {
            "-1001234567890": {
              "requireMention": true,
              "topics": { "42": { "title": "ops", "models": [{ "model": "claude-code/sonnet" }] } }
            }
          }
        }
      ]
    }
  }
}
```

`models` is an ordered sequence: the first entry leads every turn, the rest are
fallbacks tried in order when it fails. A failed attempt that already ran tools
is never failed over automatically — a rewound transcript cannot undo a command
or a sent message — so the failure comes with two buttons instead: **Continue**
picks the dead turn up where it stopped, **Restart** drops that attempt and runs
the request again from nothing. Either way the rest of the sequence takes over.
Any scope — workspace, group, topic —
may carry its own `models` sequence, and the most specific one replaces the
inherited sequence outright. `tools` narrows capabilities (`read`, `bash`,
`edit`, `write`, and `web`/`agent` on the Claude Code runtime); a workspace's
list is always intersected with the running model's. `users` and `groups` are
the allowlists — empty means deny, and pairing fills them in from the
dashboard. Values written as `"$VAR"` are read from the environment at load
time.

A scheduled prompt is a Markdown file with flat YAML frontmatter:

```markdown
---
name: daily-ingest
description: Import yesterday's inbox
schedule: "0 2 * * *"
model: claude-opus-5
---

Read yesterday's messages and update the workspace.
```

`name` and `schedule` are required. `model` may be a provider-qualified model or
an alias declared by the source. Set `enabled: false` to pause one manifest.
eleven reserves an occurrence on disk before it starts the turn, so a restart
cannot launch the same cron minute twice. A turn interrupted by a recent daemon
restart resumes from its existing transcript. Older ambiguous runs are marked
failed instead of being replayed, since their tools may already have produced
external side effects. Use `eleven schedules` to inspect jobs and
`eleven schedules run <name>` to start one manually without changing its next
cron occurrence.

eleven

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

eleven

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README