Veles

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Français · Italiano · Português (BR) · Português (PT) · Русский · العربية · हिन्दी · বাংলা · Tiếng Việt

A minimal CLI agent framework that gets smarter with every session.

Unlike chat tools that start fresh every time, Veles maintains structured project memory — insights, rules, and curated knowledge that accumulate across sessions and make the agent more useful the longer you use it. How your content is organised is pluggable: a Karpathy-style LLM wiki by default, flat notes, or no structure at all for code repos. Built clean: no god-files, no vendor lock-in, no cloud sync.

uv tool install veles-ai          # installs the `veles` command
veles init && veles run "Summarize the project architecture."
veles        # interactive REPL (just run `veles` with no subcommand)

Why Veles?

Compounding memory — Every session is distilled by the Curator into per-project memory (insights, behavioral rules, session digests in .veles/). The agent recalls relevant facts and past decisions automatically — you stop re-explaining the same context. Memory works under any content layout.

Pluggable content layouts — veles init scaffolds a Karpathy-style LLM wiki by default; --layout notes gives a flat notes directory; --layout bare adds no structure at all (ideal for code repos). Custom layout packs are a single TOML file in ~/.veles/layouts/.

Provider-agnostic routing — OpenRouter, Anthropic, OpenAI, Gemini, Ollama, llamacpp, or your claude/gemini CLI subscription. Different task types (planning, compression, insights) can route to different models.

Skills that accumulate — Reusable prompt-blocks become agent tools. Promote a skill from a project to user-global and it's available everywhere. Built-in dedup finds near-duplicate skills before they drift.

Local-first + sandboxed — No telemetry, no cloud sync. The agent sees only the active project directory. Trust ladder prompts for every sensitive tool call; pre-grant for CI.

Modular, not monolithic — Minimal core (memory, agent loop, provider protocol, tool registry). Everything else — daemon, Telegram gateway, deep research, job scheduler — is an optional, loadable module.

Quick Start

Requirements: Python 3.13+, macOS / Linux (Windows best-effort). Install uv first.

# 1. Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Install veles (the package is published as `veles-ai`; the command is `veles`)
uv tool install veles-ai
# …or from source:
#   git clone https://github.com/denisotree/veles.git && cd veles && uv tool install .

# 3. Set an API key — OpenRouter is recommended (access to all models, one key)
export OPENROUTER_API_KEY=sk-or-v1-...

# 4. Create a project
mkdir my-project && cd my-project
veles init

# 5. Talk to the agent
veles run "Read AGENTS.md and describe this project."

Open the interactive REPL instead (bare veles does the same):

veles

On first run, a setup wizard walks you through your preferred language, LLM provider, API key, default model, colour theme, and whether to initialise a project in the current directory.

Providers

Override per-run:

veles run --provider anthropic --model anthropic/claude-opus-4-8 "..."
veles run --provider ollama --model llama3.2 "..."

Store API keys in the OS keychain instead of environment variables:

veles secret set OPENROUTER_API_KEY    # prompts for value, stores in keychain

Core Workflow

Pick a content layout

veles init                  # default: Karpathy-style LLM wiki (sources/ + wiki/)
veles init --layout notes   # a single flat notes/ directory
veles init --layout bare    # no content scaffold — code repos, free-form work

The agent's own memory (insights, rules, session digests in .veles/) works identically under every layout. Custom packs are one layout.toml in ~/.veles/layouts/<name>/.

Build a knowledge base (llm-wiki layout)

veles add paper.pdf                   # read a source → write a wiki page
veles add https://example.com/post    # web pages, PDFs, plain text

veles run "What do we know about the authentication design?"
veles curate                          # explicit session → memory consolidation

The Curator runs automatically after sessions. Insight extraction catches phrases like "always prefer X" or "never do Y" and writes them as persistent project insights.

Deep research

veles research "What are the trade-offs between SQLite and PostgreSQL for this use case?"

Decomposes the question into parallel sub-questions, explores each, and synthesises a structured report.

Long-running goals

veles goal start "Migrate auth module to the new provider" --max-cost-usd 2.00
veles goal list
veles goal checkpoint <id> "Completed step 1: identified all call sites"

Scheduled jobs

veles job add --name "weekly-review" --schedule "0 9 * * 1" --prompt "Generate a weekly progress summary"
veles job list

Model Routing (Ensembles)

Route different task types to different models — set it once and forget it.

Via CLI:

veles route show                                          # current routing table
veles route set compressor anthropic/claude-haiku-4-5    # typed override
veles route reset compressor                             # back to default

Via natural language in AGENTS.md:

## Routing
Use Opus for planning and architecture decisions.
Haiku is fine for compression and insight extraction.

veles route refresh    # parse the NL hints; typed overrides always win

Skills and Modules

Skills are reusable prompt-blocks (SKILL.md) that become agent tools automatically.

veles skill add https://github.com/org/skill-repo    # install from git
veles skill add ./local-skill-dir                    # or from local path
veles skill list                                     # list with telemetry
veles skill promote my-skill                         # copy to ~/.veles/skills (global)
veles skill dedup                                    # find near-duplicates
veles skill suggest-promote --save                   # propose promotions based on usage

Modules are Python plugins that can hook into the agent lifecycle (pre_turn, post_turn, pre_tool_call, post_tool_call) and veto tool dispatches.

veles module add https://github.com/org/module-repo
veles module list

Interactive session (REPL)

veles                        # new session (bare `veles` launches the interactive REPL)
veles -c                     # continue the most recent session in this project
veles --resume <id>          # resume a specific session

Slash commands surface everything live — /status, /tokens, /context, /mode, /help — and Shift+Tab cycles modes (auto / planning / writing / goal).

Slash commands: /help · /model · /mode · /status · /tokens · /context · /wiki · /save <slug> · /history · /insights · /rules · /daemon and more.

Daemon + Telegram

Run Veles as a persistent daemon with an HTTP/WebSocket API. In a fresh project directory, veles daemon start walks you through the setup — initialize the project, enable the daemon, and connect a channel: first pick a channel type (Telegram is the only platform today, but the picker is the seam new channels register on), then fill that channel's fields (bot token, whitelist). No need to open the TUI first.

veles daemon start                        # wizard (fresh dir) → starts on 127.0.0.1:8765 (next free port if taken)
veles daemon status                       # is it running?
veles daemon list                         # daemons across all projects

Bare veles daemon opens a live control panel — a tree of project → daemons → channels. Start, stop, restart, or delete daemons, and add/remove channels (the same channel-type-first flow, key c) across every project, all from the keyboard:

The same channel wizard is also available standalone (veles channel add) on an already-running project.

API endpoints: POST /v1/runs to submit a prompt, WS /v1/runs/{id}/events to stream the response, GET /v1/sessions to list sessions. All except GET /v1/health require Authorization: Bearer <token> (mint one with veles daemon token add <name>).

Each Telegram user gets a persistent session. Use veles channel list-sessions / reset-session to manage mappings.

Multi-project

veles project list                       # registered projects
veles project switch <slug>              # print the absolute path
cd $(veles project switch <slug>)        # jump to a project

veles subproject init frontend           # create a child project
veles subproject suggest --save          # agent-detected topic clusters → proposals

Trust and Safety

Every sensitive tool call (shell execution, file writes, URL fetches) prompts:

Tool 'run_shell' wants to execute. Allow?
  [1] Once  [2] Always for this project  [3] Always everywhere  [4] Refuse

Pre-grant for CI or extended autonomous runs:

veles trust set run_shell --scope project   # pre-grant for this project
veles autopilot enable --until +2h          # temporary trust bypass (audit-logged)
veles autopilot disable

The agent sees only the active project directory — other projects, symlink escapes, and .. traversal are blocked.

Export / Import

veles export full ./backup.tar.gz        # full backup: memory, sessions, telemetry
veles export template ./template.tar.gz  # sanitised template (no sources/sessions/PII)
veles import ./backup.tar.gz --into ./new-dir

CLI Reference

Every command has --help.

Documentation

Full documentation — Diátaxis-organized (tutorials · how-to guides · reference · explanation):

• English: docs/en/index.md

Other languages: use the 🌐 switcher at the top of any documentation page.

Contributing

Contributions are very welcome — Veles is built to be extended. The core stays small (agent loop + project memory + provider protocol); almost everything else is a pluggable extension point, so adding a capability rarely means touching the core:

• Provider adapters (src/veles/adapters/) — wire up a new model backend.
• Skills — reusable prompt-blocks and tools with extends: inheritance, promotable from a project to user-global.
• Tools — typed Python the agent writes and reuses, under <project>/.veles/tools/.
• Layout packs — a single layout.toml in ~/.veles/layouts/<name>/ defines a whole content layout.
• Module hooks — observability, logging, and policy via pre_turn / post_turn hooks (src/veles/core/modules.py).
• Channels & MCP servers — new gateways and external tool sources.
• Locales — translations in src/veles/locales/.

git clone https://github.com/denisotree/veles.git && cd veles
uv sync                              # runtime + dev dependencies
uv run pytest                        # the full suite (3200+ tests, no network)
uv run ruff check src tests && uv run mypy

The codebase is deliberately decomposed — single responsibility, no god-files. Read CONTRIBUTING.md for conventions and CODE_OF_CONDUCT.md before opening a PR. Good first contributions: provider adapters, workflow skills, module hooks, and locale files.

License

Apache 2.0 with patent grant — see LICENSE and NOTICE.