Skip to content

roadmap.md

Latest commit

 

History

History
94 lines (78 loc) · 4.28 KB

roadmap.md

File metadata and controls

94 lines (78 loc) · 4.28 KB

Roadmap

Integration plan for new capabilities. Ordered by impact/effort ratio.

P0 — High Impact, Low Effort

Firecrawl CLI + Skill

  • Install firecrawl-cli globally, set FIRECRAWL_API_KEY in env
  • Build a skill teaching the agent when/how to use firecrawl commands
  • Commands available: firecrawl scrape, firecrawl crawl, firecrawl search, firecrawl map, firecrawl agent, firecrawl browser
  • Upgrades /evolve (better signal collection via firecrawl search), /absorb (crawl entire doc sites), /learn (clean scrape any page), general chat (web research via firecrawl agent)
  • No MCP server needed — CLI is callable directly from claude -p via Bash
  • Install: npx -y firecrawl-cli@latest init --all --browser
  • Free tier: 500 credits. Hobby: $16/mo (3k credits)
  • Source: https://docs.firecrawl.dev/sdks/cli

Cloudflare /crawl Skill

  • Build a skill that calls Cloudflare Browser Rendering REST API
  • Use for bulk crawling in /evolve signal collection as a free fallback
  • Free tier: 5 crawls/day, 100 pages/crawl, 10 min browser/day. Paid: $0.09/hr, 100k pages
  • AI extraction built-in (Workers AI or BYO model)
  • Sitemap-aware, robots.txt compliant, R2 caching
  • MCP option: @cloudflare/playwright-mcp for interactive browser automation
  • Source: https://developers.cloudflare.com/browser-rendering/rest-api/crawl/

P1 — Quick Wins

Vision (Image Analysis)

  • Claude Code already reads images natively via the Read tool
  • Wire Telegram photo handler (_handle_photo) to send images to claude -p
  • Use cases: screenshot analysis, diagram understanding, OCR, UI inspection
  • Cost: ~$0.004 per 1MP image (~400 extra input tokens)
  • Effort: Low — already built into Claude, just need to pipe photos through

Edge TTS (Text-to-Speech)

  • Add /speak command — convert agent response to voice message in Telegram
  • pip install edge-tts — free, 300+ neural voices, 40+ languages
  • Telegram voice messages use Opus/OGG format (need MP3 -> OGG conversion)
  • Risk: unofficial Microsoft API, could break
  • Effort: Low — pip install + ~30 lines of code
  • Source: https://github.com/rany2/edge-tts

P2 — Infrastructure

Docker Sandbox

  • Run /absorb and /evolve in isolated Docker containers
  • Prevent absorbed code from touching the host filesystem
  • Build image with Claude CLI pre-installed
  • Effort: Medium — Dockerfile + orchestration
  • Cost: Free (your own infra)
  • Best for: <5 concurrent agents on a single server

P3 — Future / Scale

E2B (Sandboxed Execution)

  • Agent-focused Firecracker microVMs, <500ms startup
  • Used by Perplexity, Manus, Hugging Face
  • SDK: sandbox.commands.run('claude -p "..."')
  • Hobby: free ($100 credit). Pro: $150/mo
  • Best for: multi-tenant isolation, productionizing
  • Source: https://e2b.dev

Daytona (Dev Environments)

  • 61.5k stars, <90ms sandbox creation, open-source
  • Snapshots (save/resume state), stateful, multi-region
  • $200 free compute included
  • Best for: long-running stateful agents, computer-use scenarios
  • Source: https://daytona.io

ElevenLabs / OpenAI TTS

Container Backend Abstraction

  • Abstract Docker/E2B/Daytona behind a common interface
  • Config toggle: sandbox.backend: docker | e2b | daytona
  • Per-pipeline sandbox policy (e.g. /absorb always sandboxed, chat runs local)

Status

Item Priority Status
Firecrawl CLI + Skill P0 Done — CLI installed, skill created. Needs firecrawl login to authenticate.
Cloudflare /crawl Skill P0 Done — skill created. Needs CF_API_TOKEN + CF_ACCOUNT_ID in .env.
Vision (photo handler) P1 Done — _handle_photo wired with cleanup, reply context, TTS.
Document handler P1 Done — _handle_document for PDFs, code files, text files.
Edge TTS / /speak P1 Done — voice.py with edge-tts + whisper.cpp STT.
Docker Sandbox P2 Planned
E2B P3 Planned
Daytona P3 Planned
ElevenLabs / OpenAI TTS P3 Planned
Container Abstraction P3 Planned