open-strix 0.1.17

Minimal autonomous agent harness with LangGraph Deep Agents

open-strix

An AI agent that talks to you over Discord.

• Memory blocks + files, all committed to Github
• Web fetch/search, files, bash/powershell, subagents
• Skills

Install uv

Install uv first:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Official install docs (alternate methods like Homebrew, pipx, winget):

• https://docs.astral.sh/uv/getting-started/installation/

Quick start (recommended)

uvx open-strix setup --home my-agent --github
cd my-agent
uv run open-strix

Run setup explicitly first (uvx open-strix setup ...), then run uv run open-strix from inside that home.

open-strix setup bootstraps the target directory with:

• state/
• skills/
• blocks/
• logs/events.jsonl
• logs/journal.jsonl
• scheduler.yaml
• config.yaml
• checkpoint.md
• pyproject.toml
• uv.lock
• .env (template)

It also runs:

• uv init --bare --python 3.11 --vcs none --no-workspace
• uv add open-strix

It also prints a CLI walkthrough with links and step-by-step setup for:

• MiniMax M2.5
• Kimi/Moonshot
• Discord bot creation + permissions
• config.yaml values

Then uv run open-strix connects to Discord if a token is present (by default DISCORD_TOKEN). Otherwise it runs in local stdin mode.

Installed mode (optional)

If you prefer a local project install instead of uvx:

uv init --python 3.11
uv add open-strix
uv run open-strix setup --home .
uv run open-strix

Install and auth gh (GitHub CLI)

If you want open-strix setup --github, install and log into gh first.

Install:

# macOS (Homebrew)
brew install gh

# Ubuntu / Debian
sudo apt install gh

# Windows (winget)
winget install --id GitHub.cli

Authenticate:

gh auth login
gh auth status

Official docs:

• https://cli.github.com/
• https://github.com/cli/cli#installation

Create a GitHub repo and set remote

open-strix auto-syncs with git after each turn, so set up a repo + remote early.

Recommended:

uvx open-strix setup --home my-agent --github

Keep this private, since agent memory and logs can contain sensitive context.

Manual fallback with GitHub CLI (gh):

cd my-agent
gh auth login
gh repo create <repo-name> --private --source=. --remote=origin
git add .
git commit -m "Initial commit"
git push -u origin HEAD

Manual fallback with GitHub web UI:

1. Create a new private empty repo on GitHub (no README, no .gitignore, no license).
2. In your project directory:

git init
git add .
git commit -m "Initial commit"
git branch -M main
git remote add origin [e-mail]:<your-user>/<repo-name>.git
git push -u origin main

If you prefer HTTPS:

git remote add origin https://github.com/<your-user>/<repo-name>.git

Check remote config:

git remote -v

Environment setup

Start from the example env file:

cp .env.example .env

Default model setup in this project expects an Anthropic-compatible endpoint:

• ANTHROPIC_API_KEY
• ANTHROPIC_BASE_URL

Discord runtime uses:

• DISCORD_TOKEN

Optional:

• DISCORD_TEST_CHANNEL_ID
• OPEN_STRIX_TEST_MODEL

Models

Default: MiniMax M2.5

This project defaults to:

• model: MiniMax-M2.5 in config.yaml
• provider prefix anthropic: internally (so the runtime uses anthropic:MiniMax-M2.5)

Use MiniMax's Anthropic-compatible endpoint in your .env:

• ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic

MiniMax docs:

• Anthropic compatibility + model IDs: https://platform.minimax.io/docs/api-reference/text-anthropic-api
• AI coding tools guide (M2.5 context): https://platform.minimax.io/docs/guides/text-ai-coding-tools

Alternative: Kimi K2.5

If you want Kimi instead of MiniMax:

1. Point Anthropic-compatible env vars at Moonshot:
   • ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
2. Set model in config.yaml to the current Kimi model ID you want.

Moonshot docs:

• Docs overview: https://platform.moonshot.ai/docs/overview
• K2 update post (links to current quick-start): https://platform.moonshot.ai/blog/posts/Kimi_API_Newsletter

Note: the Moonshot update posted on November 8, 2025 references kimi-k2-thinking and kimi-k2-thinking-turbo. If you refer to these as "K2.5", use the exact current model IDs from Moonshot docs/console.

Model config behavior

config.yaml key:

• model

Behavior:

• If model has no : (example MiniMax-M2.5), open-strix treats it as Anthropic-provider and uses anthropic:<model>.
• If model already includes provider:model (example openai:gpt-4o-mini), it is passed through unchanged.

Discord setup

Use Discord's Developer Portal web UI:

1. General Information: set app/bot name and any basic metadata.
2. Installation: set Install Link to None, then save.
3. OAuth2 -> URL Generator:
   • check bot
   • select practical bot permissions (focus on messaging/reactions/history/attachments):
     • View Channels
     • Send Messages
     • Send Messages in Threads
     • Read Message History
     • Add Reactions
     • Attach Files
4. Bot tab:
   • disable Public Bot
   • enable Message Content Intent
   • (later) set avatar/profile polish
5. Bot tab -> Reset Token:
   • copy token immediately (it may not be shown again)
   • set .env: DISCORD_TOKEN=<your_discord_bot_token>
6. Use the generated OAuth2 bot invite URL to add the bot to your server.

Reference docs for the same flow:

• Getting started (app creation + installation flow): https://docs.discord.com/developers/quick-start/getting-started
• OAuth2 scopes/install links: https://docs.discord.com/developers/topics/oauth2
• Permissions reference: https://docs.discord.com/developers/topics/permissions
• Gateway + intents reference: https://docs.discord.com/developers/events/gateway

Where this is configured in open-strix:

• Token env var name: config.yaml -> discord_token_env (default DISCORD_TOKEN)
• Actual token value: your .env
• Bot allowlist behavior: config.yaml -> always_respond_bot_ids

config.yaml tour

Default:

model: MiniMax-M2.5
journal_entries_in_prompt: 90
discord_messages_in_prompt: 10
discord_token_env: DISCORD_TOKEN
always_respond_bot_ids: []

Key meanings:

• model: model name (or provider:model)
• journal_entries_in_prompt: how many journal entries go into each prompt
• discord_messages_in_prompt: how many recent Discord messages go into each prompt
• discord_token_env: env var name to read Discord token from
• always_respond_bot_ids: bot author IDs the agent is allowed to respond to

Related files:

• scheduler.yaml: cron/time-of-day jobs
• blocks/*.yaml: memory blocks surfaced in prompt context
• checkpoint.md: returned by journal tool after a journal write
• scripts/prediction_review_log.py: helper for structured prediction-accuracy reviews
• skills/: user-editable local skills
• /.open_strix_builtin_skills/skill-creator/SKILL.md: packaged built-in skill source mounted as read-only
• /.open_strix_builtin_skills/prediction-review/SKILL.md: packaged built-in skill for prediction calibration

Runtime behavior note:

• Git sync (git add -A -> commit -> push) runs automatically after each processed turn.
• New agent homes are seeded with a twice-daily UTC scheduler job (09:00 and 21:00) for prediction-review calibration.

Personality bootstrap

Creating an agent is less about code, and a whole lot more about the time you spend talking to it. Lily Luo has a great post on forming agent personalities.

You should plan on spending time:

• Communication patterns — correct the agent to know when and how often it should use the send_message and react tools. Agents often initially find it surprising that their final message is ignored, so they need to use their tools instead.
• Talk about things you're interested in, see what the agent becomes interested in

Tests

uv run pytest -q

Discord coverage includes:

• unit tests with mocked boundaries in tests/test_discord.py
• live integration tests against real Discord in tests/test_discord_live.py

Live test env vars:

• DISCORD_TOKEN (required for live connect test)
• DISCORD_TEST_CHANNEL_ID (optional; enables live send-message test)

Safety baseline

• Agent file writes/edits are blocked outside state/.
• Reads still use repository scope.
• This is intentionally simple and should not be treated as production-ready.

License

MIT. See LICENSE.