Every tool currently spending money or holding architectural surface area, with an explicit fate. Built to answer Mishaal’s question: “What can I cancel without breaking anything?”
Verdicts: KEEP · SWAP · EVALUATE · REMOVE-WHEN-V5-DIES · REMOVE-NOW · DECOMMISSIONED
Tier 1 — Infrastructure (KEEP all)
| Tool | Monthly $ | Verdict | Why |
|---|
| Cloudflare Workers + KV + D1 + R2 + Vectorize + DOs + Cron + Workflows + Access | ~$25 (Workers Paid) + usage | KEEP | Edge-native runtime for Hermes-V3, Composio Shim Worker, custom GTM logic. Bundled plan, all in one bill. |
| Cloudflare AI Gateway | $0 | SWAP → Helicone + Langfuse | OSS observability stack gives prompt-level traces + replay; AIG is shallow. Decision pending COMPOSIO-V6 plan. |
| GitHub | $4 (Copilot) | KEEP | Code, CI, secret store. |
| Infisical | $0 (free tier) | KEEP | Canonical secret store per ~/.claude/CLAUDE.md Credential-First Mandate. |
| Hindsight | self-hosted on CF | KEEP | Long-term memory banks (mishaal, ascend-infra, kahuna). No replacement. |
Tier 2 — Migration targets
| Tool | Monthly $ | Verdict | Trigger / Notes |
|---|
V5 Worker (ascend-gateway-v5) | included in CF | REMOVE-WHEN-V5-DIES | Frozen now, deleted when last consumer (Hermes, IDEs, n8n) is repointed. No deletes during freeze. |
| Nango Cloud | $50 (Starter) | REMOVE-WHEN-V5-DIES | KEEP-TEMPORARILY. Nango writes tokens into V5’s KV. Composio holds tokens server-side and bypasses V5’s KV entirely. Kill in same PR that removes the last nango_managed:* flag. Saves $600/yr. |
| n8n Cloud | ~$50 | REMOVE-NOW (after Hermes-V3 ships) | Consumer-only since Phase 2. 83 active workflows are demo/legacy. Audit + cancel during HERMES-V3 plan. |
| VPS — ascend-agent-vps (Hetzner CPX41) | €46.49 | BLOCKED — cannot delete | RUNNING. Hindsight Docker container is live. Deletion gated on Phase F (Mem0 migration). See VPS-NUKE.md. |
| VPS — hermes server (Hetzner CPX21) | €13.99 | REMOVE-NOW (confirm) | RUNNING but no active consumers (hermes-gateway.service = old V2). Safe to delete after Mishaal confirms. |
| Mac Bridge | $0 | DECOMMISSIONED | Already off. |
Tier 3 — Composio (the new spine)
| Tool | Monthly $ | Verdict | Notes |
|---|
| Composio (Hobby → Growth) | $29 (200K calls) → $229 (2M) | KEEP — primary | New OAuth + token + MCP layer. Project ID pr_8AjnN9gdcrYt. Org ok_2IA5gRAMz7xt. Free tier 20K calls/mo to start. Multi-tenant via user_id param (one project, all clients). |
Tier 4 — LLM providers
| Provider | Use | Verdict |
|---|
| Anthropic (Claude) | User-facing IDE + claude MCP tool only | KEEP — gated |
| DeepSeek | Default cron / background workhorse | KEEP — but standardize on deepseek-v4-pro (75% discount until 2026-05-31; see ROOT_CAUSE in session memory) |
| OpenAI (gpt-4.1, o3) | Tri-judge grader | KEEP |
| xAI (Grok) | Tri-judge grader | KEEP |
| Gemini | Tri-judge grader | KEEP |
| DeepInfra / Together / Groq / Moonshot / Zhipu | Open-weight access (Qwen3-Coder, Kimi K2.6, MiniMax M2.5, GLM-5.1, gpt-oss-120B) | EVALUATE — wire one as Helicone fallback; pay-as-you-go, no commitment |
Tier 5 — Observability (decision pending)
| Tool | Monthly $ | Verdict | Notes |
|---|
| Helicone (proxy) | $0 free → $20 Pro | EVALUATE — adopt | LLM observability. Drop-in proxy in front of Anthropic/DeepSeek/OpenAI. Replaces CF AI Gateway. |
| Langfuse (OSS) | $0 self-host on CF | EVALUATE — adopt | Trace storage + eval UI. Pairs with Helicone. |
| Sentry | $0 free → $26 Team | EVALUATE | Error tracking for Hermes-V3 + Shim Worker. |
| BetterStack | $0 free → $25 | EVALUATE | Uptime + log retention beyond CF. |
| Slack (notifications) | $0 (existing workspace) | KEEP | Outbound channel, not an LLM endpoint. |
Tier 6 — AWS (KEEP — no Composio equivalent)
| Service | Use | Verdict |
|---|
| AWS Bedrock | LLM fallback for sovereignty / models not on Composio | KEEP — direct via aws4fetch |
| AWS SES | Outbound email | KEEP |
| AWS Textract | OCR for client docs | KEEP |
Tier 7 — Data providers (use through Composio after migration)
All accessed via mcp__ascend-gateway__* today. Migrate to mcp__composio__* per Phase 2 of UNPLUG-V5 trio. Catalog audit (2026-05-15) confirms 12/12 priority providers covered:
| Provider | Composio tools | OAuth managed by Composio | Verdict |
|---|
| HubSpot | 232 | yes | SWAP to Composio |
| Salesforce | 179 | yes | SWAP to Composio |
| Slack | 145 | yes | SWAP to Composio |
| QuickBooks | 105 | yes | SWAP to Composio |
| Google Analytics | 67 | yes | SWAP to Composio |
| Gmail | 61 | yes | SWAP to Composio |
| Gong | 57 | yes | SWAP to Composio (verify us-58396 region routing) |
| Apollo | 48 | API-key | SWAP to Composio |
| Google Calendar | 44 | yes | SWAP to Composio |
| SEMrush | 37 | API-key | SWAP to Composio |
| Google Ads | 9 | yes | SWAP — verify MCC routing + dev-token handling |
| Perplexity | 9 | API-key | SWAP to Composio |
| Gamma | 5 | API-key | SWAP to Composio (registered in V5 but no live consumer; rewire on Hermes-V3 cutover) |
| Google Super (unified) | 404 | OAuth2 (bundled scopes) | SWAP — DEFAULT for new connections. 1 OAuth replaces Gmail+GCal+GA4+GoogleAds + adds Drive/Sheets/Docs/Slides. Keep individual toolkits as escape hatch for compliance-sensitive clients. Verify GA4 v1beta routing + Google Ads MCC/dev-token before cutover. |
Total tool surface available via Composio: 993 tools across the 12 priority providers, +404 via googlesuper bundle, +5 via Gamma.
Tier 8 — Legacy / unused (REMOVE-NOW)
| Tool | Notes |
|---|
Old COMPOSIO_API_KEY in ~/.bash_profile | Stale key ak_wdgqHw6Yo7xQtKfpJ23x. Delete after new key in ~/.zprofile is verified end-to-end. |
Bottom-line monthly savings on full kill
- Nango: −$50 (when V5 dies)
- n8n Cloud: −$50 (when Hermes-V3 ships)
- Composio added: +$29 (Hobby) → +$229 (Growth, when calls > 200K/mo)
- Helicone Pro (optional): +$20
- Langfuse: $0 (self-host on CF)
Net monthly delta: −$71 (free Composio tier) → +$78 (Hobby + Helicone + Langfuse) → +$224 (Growth tier when usage justifies)
Tradeoff: lose 31 V5 tools (10 of which Composio replaces 1:1; 21 require Shim Worker for AWS / agent state / RAG / cron / platform tools).
Open evaluations (not blocking)
- CF AI Gateway → Helicone+Langfuse swap — needs ADR before execution.
- Sentry vs CF Logpush+R2 Iceberg for error storage — duplicate spend if both adopted.
- BetterStack vs CF Logpush — same question.
- Open-weight model fallback chain ordering — needs head-to-head benchmark on Ascend GTM eval set.