Instant context routing for AI agents. A cheap model watches your conversation and automatically injects the right docs and skills before your agent responds — so it never forgets to read what it needs.

You document everything. You have an index with summaries. You have read_when hints everywhere. You have reminders in CLAUDE.md. And still, the agent doesn't read the right docs before coding.

The problem isn't missing documentation — it's that static injection doesn't adapt to what you're actually asking right now.

On every user message, Reflex:

1. Reads the last 10 conversation entries
2. Looks at your registry of available docs and skills
3. Uses a fast model (gpt-5.2 by default, any OpenAI-compatible endpoint) to decide what's relevant
4. Injects a directive to read relevant docs and invoke relevant skills
5. Tracks what's already been injected — no duplicates within a session

Your agent sees the injection as part of the current message — not buried in a system prompt.

1. Install the reflex binary

curl -sL https://raw.githubusercontent.com/markmdev/reflex/master/install.sh | sh

Or with Go:

go install github.com/markmdev/reflex@latest

2. Set your API key

reflex config set api-key sk-your-openai-key

Or use an environment variable:

export OPENAI_API_KEY=sk-your-key

3. Install the Claude Code plugin

/plugin marketplace add markmdev/claude-plugins
/plugin install reflex@markmdev

That's it. No registry to maintain — Reflex discovers everything automatically.

Skills — add name and description frontmatter to .claude/skills/*/SKILL.md:

---
name: planning
description: "Create implementation plans. Use for new features, refactoring, architecture."
---

Docs — add summary and read_when frontmatter to any .md file:

---
summary: "OAuth implementation details and gotchas"
read_when:
  - authentication
  - OAuth
  - login
---

Reflex discovers both automatically. No list to maintain.

Reflex is a framework-agnostic CLI. Hooks call it; you can also call it directly.

echo '{
  "messages": [{"type": "user", "text": "help me set up OAuth"}],
  "registry": {
    "docs": [{"path": "auth.md", "summary": "OAuth guide", "read_when": ["OAuth"]}],
    "skills": []
  },
  "session": {"docs_read": [], "skills_used": []}
}' | reflex route

# Output:
{"docs":["auth.md"],"skills":[]}

View recent routing decisions:

reflex logs

Any OpenAI-compatible API works. Default: OpenAI gpt-5.2 via Responses API.

Config lives at ~/.config/reflex/config.yaml:

provider:
  base_url: https://api.openai.com/v1
  model: gpt-5.2
  responses_api: true

Switch providers by changing those lines:

# Kimi K2.5
provider:
  base_url: https://api.moonshot.ai/v1
  api_key_env: MOONSHOT_API_KEY
  model: kimi-k2.5

Or use the CLI:

reflex config set model gpt-4o-mini
reflex config set base-url https://api.openai.com/v1
reflex config show

Input (stdin JSON):

{
  "messages": [{"type": "user", "text": "..."}],
  "registry": {
    "docs": [{"path": "file.md", "summary": "...", "read_when": ["keyword"]}],
    "skills": [{"name": "planning", "description": "..."}]
  },
  "session": {"docs_read": ["already-read.md"], "skills_used": []},
  "metadata": {}
}

Output (stdout JSON):

{"docs": ["file.md"], "skills": []}

Returns {"docs": [], "skills": []} when nothing is needed. Errors go to stderr. Always exits 0 — never blocks your agent.