nerve

---

Nerve is a self-hosted runtime for AI agents, built around the Claude Agent SDK. It gives agents everything they need to be useful long-term: persistent memory, scheduled execution, task management, learnable skills, and channels to reach you through — web UI, Telegram, Slack, or autonomous cron jobs.

Ship a personal assistant that develops a personality, remembers your preferences, and manages your inbox. Or deploy a worker agent that monitors your CI, reviews PRs, and fixes flaky tests — all plan-driven with human approval. Same engine, different mission.

Quick Start

curl -fsSL https://raw.githubusercontent.com/ClickHouse/nerve/main/install.sh | bash

The installer puts the dependencies in place (Python, Node.js, uv), clones the repo and runs nerve init. That asks three questions in the terminal — server or Docker, where the files live, which port and address — then starts Nerve and prints the address of the setup page with a one-time setup token. Everything else (your account, personal assistant or worker, the model provider, channels, memory, scheduled jobs) is decided in the browser. See the Setup Guide.

Prerequisites: Git, Python 3.13+, uv, Node.js 22.12+

git clone https://github.com/ClickHouse/nerve.git && cd nerve
uv sync                      # creates .venv from uv.lock
source .venv/bin/activate
nerve init                   # builds the web UI, starts Nerve, prints the setup page + token

Prefer to answer everything in the terminal? nerve init --terminal is the full wizard; it does not start Nerve, so follow it with nerve start. For hosts with no terminal at all, see unattended installation.

Upgrading

nerve upgrade
nerve restart

Runs git pull, reinstalls Python deps, and rebuilds the web UI in one shot. Use --no-pull, --no-deps, or --no-frontend to skip steps. For Docker deployments, rebuild the image instead: git pull && docker compose build && docker compose up -d.

No API key? Use your Claude subscription instead

Two Modes

🧑 Personal Mode

A full-featured assistant for one human. Syncs your email, remembers your preferences, develops personality over time. Comes with a web UI, Telegram bot, memory, cron jobs, and notifications out of the box.

"You assist one person over the long run." — from the Personal SOUL template

• Workspace files define personality (SOUL.md), identity (IDENTITY.md), and user context (USER.md)
• Memory categories are life-oriented: relationships, finances, travel, health, work
• Default crons: inbox processor, task planner, memory maintenance
• The agent develops opinions, follows up on past conversations, and gets better the longer you use it

🔧 Worker Mode

A task-focused autonomous agent for teams or programmatic deployment. Monitors something, proposes fixes, implements after approval. Plan-driven with a full audit trail.

"You are a specialist with one mission, set out in TASK.md, and an explicit quality bar." — from the Worker SOUL template

• Self-configuring: tell it what to work on in its onboarding conversation, or give it a plain-English task description at install time, and it researches, writes its own TASK.md, creates skills, sets up cron jobs, and starts working
• Memory categories are operational: patterns, procedures, decisions, approvals, infrastructure
• No IDENTITY.md or USER.md — workers execute a mission, not build a relationship
• Built for reliability: plan before acting, verify after executing, log everything

The mode is chosen on the setup page: run nerve init, open the page it prints, and pick Worker on the Mode step. When you finish setup, the agent opens its onboarding conversation: it starts from the note you left on the Mode step (or asks what to work on), studies the repositories and sources in the background, then asks you the few things only you know, proposes the setup as a plan you approve, edit or send back, and applies it. The same choice in a terminal is nerve init --terminal; unattended installs set NERVE_MODE=worker and NERVE_TASK. Without the browser, a worker with a task description sets itself up at its first start, in one session without a person, and notifies you when it is done. With an empty task description, it waits for its onboarding conversation in the web UI.

# Spin up a worker that monitors CI and fixes flaky tests, with no browser
NERVE_MODE=worker NERVE_TASK="Monitor CI for flaky tests and propose fixes" \
ANTHROPIC_API_KEY=sk-ant-REPLACE NERVE_PASSWORD=REPLACE \
nerve init --non-interactive && nerve start

Why Nerve?

🧠 Memory that persists	Dual-layer memory: curated hot memory loaded every session + semantic search over your entire history. Agents remember context, preferences, and lessons learned — permanently.
🔧 Self-evolving skills	Agents watch for repeated workflows and propose reusable skills automatically. Skills are plain Markdown — agents read, create, and improve their own procedures over time.
📋 Autonomous task planning	A background planner explores open tasks, researches codebases, and proposes implementation plans — without being asked. You review and approve; the agent executes.
📡 Multi-source awareness	Gmail, GitHub, and Telegram flow into a unified inbox with Kafka-style cursors. Multiple consumers process the same data independently — inbox triage, digest generation, task extraction.
⚡ Single process, zero ops	No Docker required, no message queues, no service mesh. FastAPI + uvicorn + asyncio. Runs on a Raspberry Pi.
🔑 No API key required	Optional OAuth proxy routes through your Claude Max/Pro subscription. Zero API costs.
🔄 Session persistence	Sessions survive server restarts. Fork conversations, resume from any point, run up to 4 concurrent agent sessions.
🛡️ Human-in-the-loop	Agents can pause mid-turn to ask questions via any channel. Plans require approval before execution. Destructive actions need explicit consent.

Channels

Every channel shares the same agent engine, memory, and tools.

💬 Web UI

Full React + Vite + TailwindCSS frontend served by FastAPI. Real-time streaming over WebSocket.

• Chat with session history, branching, and resume
• Setup wizard: claim the instance, choose a mode, connect a model, configure channels, sources and jobs — re-enterable from the Setup entry
• Inline diff viewer for file changes (no git required)
• Task manager, memory browser, skill editor
• Source inbox, notification center, cron logs
• Plan review with approval/decline/revision workflow
• JWT auth, self-signed HTTPS via mkcert

🤖 Telegram Bot

Powered by python-telegram-bot v21+.

• Streaming responses via edit-in-place
• Inline keyboard buttons for interactive tool questions
• /reply command for free-text answers
• Configurable DM policy (open or pairing)

💬 Slack Bot

Reach your agent where your team already works — a direct message, or a channel you have invited it to. Nothing to expose: it connects outwards, so it runs behind NAT with no public URL.

• Answers appear as they are written, and stay in the thread you started
• Each thread is its own conversation, so several people can work in one channel
• In a channel it stays quiet until you @mention it
• Reply to its questions by pressing a button, or send it a file to read
• You decide who may talk to it and where — by person, by channel, or both

⏰ Cron Jobs

Scheduled AI sessions via APScheduler. Three session modes:

Mode	Behavior
Isolated	Fresh session per run. No prior context. Best for briefings, cleanup, reports.
Persistent	Accumulates context across runs. Optional rotation to manage token usage. Reminder mode for lightweight follow-ups.
Main	Runs inside the user's primary conversation — full context access.

Built-in crons (personal mode): skill-extractor (12h), skill-reviser (weekly), inbox-processor (30min), task-planner (4h), machine-maintenance (weekly). Worker mode ships with skill-reviser, skill-extractor and machine-maintenance on and task-planner off; onboarding enables task-planner when the worker should plan through the approval loop, and configures the worker's own crons.

📡 Source Sync

Cursor-based data ingestion pipeline. Each source runs as an independent APScheduler job.

• Telegram — Telethon user-account API (updates.getDifference)
• Gmail — gog CLI with 2-step fetch and LLM condensation
• GitHub Notifications — gh api enriched with PR/issue content
• GitHub Events — Pushes, PRs, reviews, comments

Multiple independent consumers read the same inbox at their own pace. Content is prefixed with untrusted-data warnings to prevent prompt injection.

Core Systems

🧠 Memory

Two layers, one seamless experience.

L1 — Hot Memory (MEMORY.md) Curated facts injected into every system prompt. Active projects, current deadlines, operational lessons. Tagged with dates, automatically evicted when stale.

L2 — Deep Memory (memU) Semantic search over everything — conversations, facts, preferences, events. SQLite-persisted. Uses vector embeddings when an OpenAI key is configured, or LLM-based ranking with Anthropic models only.

• Four memory types: profile, event, knowledge, behavior
• Automatic conversation indexing on session close
• Pre-recall: relevant memories delivered with the first message when sessions start (kept out of the system prompt so it stays cacheable across sessions)
• 3-level quality filtering prevents generic facts from polluting memory
• Semantic deduplication (cosine similarity 0.85 threshold)
• Full audit log of all mutations

Memory categories adapt to mode — personal agents track relationships, health, and finances; workers track patterns, procedures, and approvals.

📋 Tasks

Markdown files backed by SQLite FTS5 full-text search.

• Statuses: pending → in_progress → done / deferred
• Duplicate detection: exact URL match + fuzzy FTS fallback
• Escalation reminders: deadline → +30min follow-up → +2h URGENT (respects quiet hours)
• Full markdown editor in the web UI

🔧 Skills

The agent's procedural knowledge. Pure Markdown files that agents read, create, and update themselves.

workspace/skills/
└── my-skill/
    ├── SKILL.md          # Instructions (YAML frontmatter + markdown)
    ├── references/       # On-demand documentation
    ├── scripts/          # Executable code (30s timeout)
    └── assets/           # Supporting files

• Progressive disclosure: only name + description in system prompt; full content loaded on demand
• skill-extractor cron proposes new skills from repeated workflows
• skill-reviser cron reviews existing skills for accuracy
• Usage statistics tracked per skill
• Platform skills: an operator or a hosting platform can add read-only skills in skills/ in the config directory (details)

📐 Plans

Proactive task planning with human-in-the-loop approval.

1. Planner cron picks open tasks autonomously
2. Agent explores codebase, researches approaches, proposes implementation plan
3. User reviews → approves / declines / requests revision
4. Approved plans spawn implementation sessions automatically

Revisions happen in the same persistent planner session — full context preserved.

🔔 Notifications

Async communication between agent and human, delivered to the web UI, Telegram, and Slack. notifications.channels chooses which of them; all three are on by default and a transport that is off is skipped.

• notify — Fire-and-forget alerts (status updates, completions, reminders)
• ask_user — Questions with predefined options, rendered as buttons
• Priority levels: urgent, high, normal, low

Architecture

nerve (single Python process)
│
├── Gateway (FastAPI)
│   ├── REST API — sessions, tasks, memory, plans, sources, diagnostics
│   ├── WebSocket — real-time streaming, answer routing
│   └── Static files — serves the React web UI
│
├── Agent Engine (Claude Agent SDK)
│   ├── ~30 custom MCP tools (tasks, memory, skills, sources, notifications)
│   ├── Interactive mid-turn pausing (AskUserQuestion, plan approval)
│   ├── File snapshot hooks → session-scoped diffs without git
│   ├── Session lifecycle: created → active → idle → stopped → archived
│   └── Up to 4 concurrent sessions (configurable)
│
├── Channels
│   ├── Web — passive WebSocket channel
│   ├── Telegram — bot with streaming + inline keyboards
│   └── Slack — bot with streaming + buttons, per-thread sessions
│
├── Cron (APScheduler)
│   ├── AI jobs (isolated / persistent / main session modes)
│   └── Source sync jobs (Telegram, Gmail, GitHub)
│
├── Memory
│   ├── File-based hot memory (L1)
│   └── memU semantic index (L2, SQLite + embeddings)
│
├── Tasks — Markdown + SQLite FTS5
├── Skills — Filesystem + DB index + usage tracking
├── Plans — SQLite with approval workflow
├── Notifications — Multi-channel delivery
└── Proxy (optional) — Claude OAuth, no API key needed

Configuration

Three layers, lowest precedence first:

• <workspace>/config/settings.yaml — shareable settings, git-tracked with the workspace (this is the layer you review in a PR and sync between machines)
• config.yaml — machine-local settings (gitignored)
• config.local.yaml — secrets and personal overrides (gitignored)

See docs/config.md for all options.

Requirements

• Python 3.13+
• uv (recommended package manager)
• Node.js 22.12+ (for web UI build)
• Claude Code CLI (bundled with claude-agent-sdk)
• Anthropic API key or Claude subscription via CLIProxyAPI proxy
• Optional: OpenAI API key (for vector-based memory search — without it, LLM-based recall is used), Telegram bot token, gog CLI, gh CLI

Documentation

Doc	Description
Architecture	System overview, modes, deployment, data flow
Setup	Installation, the browser setup wizard, the setup token, Docker, HTTPS, systemd
Config	All config options with descriptions and defaults
Accounts	Local accounts, actor identity, the first-start bootstrap and the signing secret
Worker Guide	Worker mode setup, onboarding, plan-approve loop
Personal Best Practices	Running a personal assistant well: brain files, memory tiers, approval classes, turn shape, safety hygiene
Worker Best Practices	Running an autonomous worker well: TASK.md anatomy, plan-approve-execute, task state machine, CI results comments, escalation
API Reference	REST API and WebSocket protocol
SDK Sessions	Session lifecycle, resume, forking
Sources	Sync engine, cursors, consumer model

Testing

pytest tests/ -v