devrel-swarm 0.2.8

A 13-agent CLI that runs DevRel, sales, and marketing on your repo. BYO Anthropic key.

devrel-swarm

A developer-first CLI for AI-powered DevRel, sales, and marketing.

devrel-swarm is a pipx-installable command-line tool that runs a 13-agent system against any project — community triage, social listening, theme extraction, growth experiments, content production, video tutorials, documentation, competitive intel, sales outreach, brand-consistent campaigns, and post-publish content performance analysis. Operates on a project repo the way git, npm, and cargo do.

Every piece of content the system produces flows through an 8-stage editorial pipeline (developmental edit → line edit → copy edit → anti-slop → reader-persona test → readability check → brand audit) so output reads like senior-editor work, not generic AI prose.

Every deliverable in this repository was produced by the agent system itself.

---

Quick start

pipx install devrel-swarm

cd /path/to/your/project
devrel init --name myproject --url https://myproject.dev --github-repo me/myproject

# edit .devrel/voice.md, .devrel/style.md, .devrel/slop-blocklist.md

export ANTHROPIC_API_KEY=sk-ant-...
devrel doctor                                  # check env + scaffold
devrel content draft "tutorial on feature flags" --type tutorial
devrel run                                     # full weekly pipeline

After devrel init, your repo has a .devrel/ directory with:

.devrel/
  config.toml          # product identity, model selection, budget caps
  voice.md             # tone profile + sample passages    (commit)
  style.md             # house style + per-content targets (commit)
  slop-blocklist.md    # banned phrases                    (commit)
  kb/                  # knowledge base, TF-IDF indexed
  deliverables/        # generated outputs
  state.db             # SQLite: jobs, costs, checkpoints

The four committed files (config.toml, voice.md, style.md, slop-blocklist.md) encode the editorial contract. Diff them like any other source.

---

Commands

# Bootstrap & health
devrel init                      bootstrap .devrel/ in cwd
devrel doctor [--json]           check env, API keys, KB freshness
devrel cost [--month YYYY-MM]    token + USD report from state.db

# Pipelines
devrel run                       full weekly cycle
devrel run --health              health check only (Watchdog)
devrel run --agent NAME --task T run a single agent ad-hoc

# DevRel
devrel triage [--days N]         GitHub issue triage (Sage)
devrel listen [--platforms ...]  Reddit / HN / X (Echo)
devrel synthesize                theme extraction (Iris)
devrel experiment HYPOTHESIS     A/B + power analysis (Nova)

# Content
devrel content draft PROMPT      revision-looped + 5-lever quality (Kai)
devrel content audit FILE        run quality pipeline on existing draft
devrel content slop FILE         run only the anti-slop pass
devrel docs build                AST-based docs (Dex)
devrel video record SCRIPT       screen-recorded tutorial (Vox)

# Sales
devrel intel COMPETITOR
devrel sales outreach COMPANY
devrel sales battlecard COMPETITOR
devrel sales sequence CAMPAIGN

# Marketing
devrel marketing blog TOPIC
devrel marketing landing TOPIC
devrel marketing social TOPIC
devrel marketing campaign BRIEF

# Knowledge base
devrel kb add URL [--category C]
devrel kb list
devrel kb refresh

# Config & schedule
devrel config get KEY
devrel config set KEY VALUE
devrel schedule install | list | remove

# Outputs
devrel deliverables list
devrel deliverables show NAME

# Analytics (Argus)
devrel analytics report [--since 7d] [--push] [--push-on-partial]
devrel analytics history CONTENT_ID
devrel analytics diff PERIOD_A PERIOD_B
devrel analytics calibration
devrel analytics summary [--root PATH]

Global flags on most verbs: --json (machine-readable output) and --quiet.

---

Editorial quality pipeline

Every content-producing run (devrel content draft, devrel content audit, plus internal calls from marketing, sales, kb-driven tutorials) flows through 8 stages:

1. Generate          KB-grounded; voice.md + style.md in prompt
2. Developmental     critique+revise (structure, argument, hook)
3. Line edit         critique+revise (rhythm, voice fidelity)
4. Copy edit         critique+revise (grammar, code, consistency)
5. Anti-slop         regex blocklist + LLM lint; force-rewrite on hit;
                       second failure aborts loud with a phrase report
6. Reader persona    "skeptical senior backend dev" scores 1-10
7. Readability       Flesch + sentence variance + jargon density
                       checked against per-content-type targets
8. Brand audit       Sentinel (existing 6-dimension audit)

→ deliverables/ + revision-trace.json (every stage's score + diff)

Stages 5-7 use Haiku for cost; stages 2-4 use Sonnet. Total cost ≈ 2.5-4× a single revision loop, with prompt caching pulling toward the lower bound. BudgetGate guardrails (configurable in .devrel/config.toml) track spend; devrel cost --month YYYY-MM reports it.

---

How it works internally

Hub-and-spoke with 13 agents. Atlas orchestrates; specialists execute across three pipelines.

Atlas (Orchestrator)
├── Health: Watchdog (pre-flight) + Sentinel (pre-publish brand audit) + Argus (post-publish performance analyst)
├── DevRel: Sage, Echo, Iris, Nova, Kai, Vox, Dex
└── Sales:  Rex, Pax, Mox

The weekly cycle (driven by devrel run):

Stage 0: Watchdog         (health + budget check)
Stage 1: Sage + Echo + Dex     parallel
Stage 2: Rex + Iris            parallel
Stage 3: Nova + Kai            parallel (Kai routes through quality pipeline)
Stage 4: Vox
Stage 5: Sentinel              brand audit
Stage 5b: Argus                post-publish content performance analysis
Stage 6: Instantly sync, OKR compilation, Sheets publish, digest

Argus is config-gated by [orchestration].analytics_in_run (default true); set to false to skip the stage. Standalone use via devrel analytics report is unaffected.

The Atlas.delegate() API also dispatches single-agent tasks, which is what every non-run verb wraps. So devrel triage is Atlas.delegate("sage", "Triage GitHub issues from the last 7 days") — the agents never appear in the public CLI surface, only the verbs.

---

Configuration

.devrel/config.toml example:

[project]
name = "openclaw"
url = "https://openclaw.ai"
github_repo = "openclaw/openclaw"

[model]
default = "claude-sonnet-4-6"
cheap = "claude-haiku-4-5-20251001"
opus_opt_in = true

[budget]
monthly_usd = 100.0
warn_at_pct = 80

Edit with devrel config set <key> <value> or directly in your editor.

Environment variables

Variable	Required	Used by
ANTHROPIC_API_KEY	yes	every LLM-using verb
GITHUB_TOKEN	for triage	Sage
FIRECRAWL_API_KEY	for KB harvest + intel	kb add, Echo, Rex
BRAVE_API_KEY	optional fallback	search
INSTANTLY_API_KEY	for cold-email sync	Pax, Mox
APOLLO_API_KEY	for lead enrichment	Rex, Pax
TELEGRAM_BOT_TOKEN	for digests	Atlas pipeline
EMAIL_SENDER / EMAIL_PASSWORD	for digests	Atlas pipeline
OPENAI_API_KEY	for video TTS	Vox

.env files at the project root are loaded automatically. Cross-project shared keys can live at ~/.devrel/secrets.env.

---

Retargeting to another product

cd /path/to/other-project
devrel init --name otherproduct --url https://otherproduct.dev --github-repo owner/otherproduct
devrel kb add https://otherproduct.dev/docs --category docs
# edit voice.md / style.md / slop-blocklist.md to match the other product's voice
devrel doctor
devrel run

The agent system is product-agnostic. Per-project config + KB + voice files do all the targeting.

---

Documentation

The user-facing docs live in docs/:

• docs/quickstart.md — bootstrap a project and run your first weekly cycle in 5 minutes
• docs/agents/argus.md — content performance analyst, the 13th agent
• docs/cli/analytics.md — full reference for the devrel analytics subgroup
• docs/cookbook.md — common recipes (calibration, weekly cron, multi-project rollups)

Internal docs (architecture specs, implementation plans) live in docs/superpowers/.

---

Tech stack

Component	Choice
Language	Python 3.12+
CLI framework	Typer + Rich
Agent SDK	Claude Agent SDK
HTTP	httpx (async)
Default model	Claude Sonnet 4.6 (Haiku for cheap quality stages, Opus opt-in)
Stats	scipy (power analysis, Bayesian eval)
Video	Playwright + FFmpeg + OpenAI TTS
Storage	SQLite per project (.devrel/state.db)
Tests	pytest + pytest-asyncio + respx

---

Author

Daria Dovzhikova — DevTools Growth Strategist & AI Agent Builder

• 12+ years in DevTools (JetBrains, Huawei, Lightrun, Odigos)
• dariadovzhikova.com

---

MIT License