The AI agent that runs your social media. Not a dashboard. An autonomous system.
Built in 2 weeks by one engineer with Claude Code + Codex CLI.
For the production version, contact me.
Say "search for AI trends, then make a post for X, LinkedIn, and Threads" and the agent:
- Searches the web for current trends
- Recalls your X training (style guide + corrections + writing patterns)
- Generates an X draft
- Recalls your LinkedIn training
- Generates a LinkedIn draft
- Generates a Threads draft
- Responds (with a summary, if applies)
One message. Six tool calls. Three platform-specific drafts. All in your trained voice.
Social media tools charge $48-399/mo for calendars and stateless GPT wrappers. They forget everything between sessions.
| Tool | Monthly Cost | Memory? | Self-Hosted? | Autonomous? |
|---|---|---|---|---|
| Hootsuite | $199-399/user | No | No | No |
| Sprout Social | $199-399/seat | No | No | No |
| Later | $18-82 | No | No | No |
| Rella | $24-48 | Basic | No | No |
| Rebel Forge | $0 | Per-platform voice memory | Yes | Yes |
The agent chains tools autonomously. Up to 8 steps per turn. It decides what to call, in what order, and when to stop.
| Tool | Purpose |
|---|---|
recall_training |
Load platform-specific voice, corrections, and style before generating |
generate_drafts |
Create drafts with auto-approve and auto-publish flags |
web_search |
Search the web for trends, news, context |
generate_image |
Generate images via fal.ai or ComfyUI |
publish_draft |
Publish to any platform (platform-matched draft selection) |
approve_draft |
Approve content for publishing |
run_heartbeat |
Trigger full Scout > Analyst > Creator cycle |
update_brand |
Update voice, audience, goals |
setup_platform |
Generate bio, handle, starter posts |
query_drafts |
Query your drafts database |
save_onboarding |
Save brand profile from onboarding |
Real chains observed in production:
web_search > generate_drafts (2 tools)
recall_training > generate_drafts (2 tools)
recall_training > web_search > generate_drafts (3 tools)
web_search > generate_drafts > recall_training > generate_drafts (4 tools)
The agent is resilient to mid-chain failures. If any step fails (API timeout, provider error, rate limit), the agent re-spins the failed step and continues from where it left off. No manual intervention, no lost progress — the chain completes even when a middle step is interrupted.
The agent doesn't just remember your brand. It remembers how you sound on each platform.
General Voice sets the baseline ("No fluff. Write like a builder."). Per-platform styles override it ("X: max 2 sentences. LinkedIn: 5 paragraphs with a question."). The agent recalls the right combination before every generation.
| Layer | What It Does | Stored In |
|---|---|---|
| General Voice | Base rules for all platforms | platform_styles (platform=general) |
| Platform Style Guide | Per-platform tone override | platform_styles per platform |
| User Corrections | Original vs. edited samples with ratings | corrections table |
| Style Learning | Patterns from your real published posts | Learned from fetched post data |
Same prompt, different platform, different output. The X draft is 84 characters. The LinkedIn draft is 730.
Fetch your real posts from any connected platform. Sort by engagement, views, likes, or date. Hit "Learn Style" and the agent absorbs your actual writing patterns — per platform.
The agent uses this when recall_training fires: your corrections, your style guide, and your real writing patterns all load before content generation.
┌─────────────────────────────────────────────────────┐
│ REBEL FORGE │
├──────────┬──────────────┬──────────────┬────────────┤
│ Frontend │ Backend │ Worker │ Database │
│ Next.js │ FastAPI │ Heartbeat │ PostgreSQL │
│ :3000 │ :8080 │ + Jobs │ :5432 │
└──────────┴──────┬───────┴─────┬────────┴────────────┘
│ │
┌────────┴────────┐ │
│ LLM Provider │ │
│ (hot-swap) │ │
├─────────────────┤ │
│ vLLM (local) │ │
│ Codex CLI │ │
│ OpenRouter │ │
└─────────────────┘ │
┌────┴─────────────┐
│ Image Provider │
│ (auto-fallback) │
├──────────────────┤
│ ComfyUI (local) │
│ fal.ai (cloud) │
└──────────────────┘
LLM and image providers are hot-swappable from settings. ComfyUI down? fal.ai takes over automatically.
Both providers run on local hardware — no cloud bills, no rate limits, no data leaving your machine.
| Service | Repo | What It Does |
|---|---|---|
 |
hec-ovi/vllm-gpt | GPT-OSS 20B/120B on AMD Strix Halo via ROCm — OpenAI-compatible /v1/responses |
 |
hec-ovi/comfyui-strix-docker | FLUX / Stable Diffusion on AMD RDNA 3.5 — verified ROCm Docker setup |
Live, tested, working. The agent publishes from chat with one command.
| Platform | Text | Images | Auto-publish | Live |
|---|---|---|---|---|
 |
yes | -- | yes | yes |
 |
yes | -- | yes | yes |
 |
yes | -- | yes | yes |
 |
yes | yes | yes | yes |
 |
yes | -- | yes | yes |
Masonry layout. Platform icons. Status colors. Inline editing with character limits (280 for X). Approve > Publish workflow with edit-reverts-to-draft safety. Published posts show live permalinks.
Three agents on an autonomous loop:
Scout → web search for trends
Analyst → reviews past performance
Creator → drafts content in your trained voice
Runs on a configurable interval. You approve or let it auto-publish.
65+ endpoints. Full OpenAPI docs at localhost:8080/docs.
POST /v1/chat — Agentic chat with 11 tools + multi-step tool loop
POST /v1/drafts/generate — Generate platform-specific content
POST /v1/drafts/{id}/publish — Publish (platform-matched draft selection)
POST /v1/training/feedback — Submit voice corrections with rating
PUT /v1/training/platform-styles/{p} — Set general or per-platform style guides
GET /v1/fetch-posts/{platform} — Fetch your posts with engagement metrics
POST /v1/training/style-learn — Learn voice patterns from real posts
GET /v1/training/style-learn/{p} — Get learned style data for a platform
POST /v1/heartbeat/trigger — Trigger autonomous agent cycle
This project is at ~60% toward production. What you see here already works — agentic tool chains, per-platform voice memory, five-platform publishing, error recovery, local inference. Built in 2 weeks by one engineer.
Looking for someone who builds complex agentic systems, autonomous tooling, and production AI pipelines? That's what I do.
Built by Hector Oviedo
Engineered with AI.