Technical dossier · FromDonna research

IBM stack for Donna

What IBM can actually sell a small agentic AI startup building a multi-user personal agent (Hermes-in-E2B sandboxes, Cloudflare Worker gateway, external LLMs, Composio tools). Concise. Primary docs. No brochure paste.

Audience founders / infra Stack today E2B · CF Workers · Grok/Codex · Composio Research Exa + IBM docs + X · 2026-07 Status GA-focused, beta noted

Must-try

Credits, OpenAI-compat gateway, Granite, Code Engine free tier

Maybe later

Orchestrate, BeeAI, vectors, governance, Postgres

Skip (for now)

IKS bare metal, Satellite, enterprise-only control planes

OpenAI-compat API

Yes

watsonx model gateway · REST + OpenAI SDK

Ranked offers for Donna

Top picks ordered by expected ROI for a Telegram-first personal agent startup — not by enterprise marketing weight. Labels: must-try maybe skip

Startup with IBM — Cloud credits

Up to ~$120k/yr (Builder $1k/mo or Premium $10k/mo × 12). Free runway on watsonx + Code Engine + DBs if eligible (<5y, <$1M rev). Apply before paying for IBM Cloud.

must-trycomplement

watsonx.ai model gateway (OpenAI-compatible)

Unified OpenAI-compatible chat/completions routing to Granite + third-party models. Critical for swapping Donna’s LLM proxy without rewriting clients. Free Lite: ~300k tokens + 20 CUH/mo.

must-trycomplement

IBM Granite models (3.x / 4.x)

Apache 2.0 open weights; strong tool-calling & code; cheap hosted inference (e.g. granite-4-h-small ~$0.06/$0.25 per 1M in/out). Good secondary / cost-control model beside Grok.

must-trycomplement

IBM Cloud Code Engine

Serverless containers + jobs + functions, scale-to-zero, free monthly allotment. Candidate for long-running agent jobs / batch — not a CF Workers edge replace, not an E2B sandbox.

must-trycomplement

BeeAI Framework (Linux Foundation / IBM-origin)

Open multi-agent framework with native MCP & A2A, provider-agnostic (incl. watsonx). Useful if Donna grows multi-agent internals — optional, not required vs current Hermes.

maybecomplement

Databases for PostgreSQL (+ Redis)

Managed memory/session store if you ever colocate state on IBM Cloud. Only if you already use credits there — not better than Neon/Upstash/CF D1 for CF-native Donna.

maybecomplement

watsonx.data + Milvus / OpenSearch vectors

Lakehouse + integrated vector (Milvus) + OpenSearch hybrid search for enterprise RAG. Overkill for personal agent memory until B2B multi-tenant knowledge bases.

maybeignore (now)

watsonx Orchestrate + ADK / MCP

Enterprise agent control plane (catalog, multi-agent orchestration, governance). Different product surface than personal Telegram agents. Explore for enterprise GTM / Partner Agent Connect later.

maybeignore (now)

watsonx.governance + agent observability

Model risk, security metrics, agent monitoring. Valuable for regulated B2B tenants; heavy for consumer personal agents today.

maybecomplement

IBM Cloud Secrets Manager + IAM

Solid secrets/IAM if you land on IBM infra. CF Secrets / Wrangler secrets already cover Workers; only needed with Code Engine/VSI footprint.

maybecomplement

IBM Cloud Kubernetes Service / OpenShift

Full cluster ops. Donna deliberately avoids cluster babysitting (E2B + Workers). Skip unless enterprise customer mandates IBM K8s.

skipignore

Virtual Servers / bare metal / Satellite

Classic VPS & edge/on-prem. No advantage vs Hetzner/Fly/E2B for sandbox agents. SoftLayer-legacy UX is a known startup footgun — prefer modern VPC products only if forced.

skipignore

Apply Startup credits → watsonx.ai free trial Code Engine console

Research buckets

Seven required coverage areas. Each product: what it is, why it matters for Donna, pricing signal (no invented figures), official URL, fit vs current stack.

01 · Inference

Foundation models & APIs

watsonx.ai must-try complement

What: End-to-end AI studio: foundation models, Prompt Lab, Tuning Studio, RAG, AgentLab, deployable inference.
Donna relevance: Hosted inference + prompt tooling. Model gateway is the integration point for Donna’s LLM proxy.
Pricing signal: Free toolbox: ~300k tokens/mo + 20 CUH. Essentials PAYG from $0; Standard from ~$1110/mo. Token rates per model (see pricing page).
URL: ibm.com/products/watsonx-ai · pricing

IBM Granite (3.x / 4.x) must-try complement

What: IBM’s open foundation model family (language, code, vision, guardian). Apache 2.0 weights; dense + MoE variants; improved tool calling in 4.1 (3B/8B/30B).
Donna relevance: Cost-efficient tool-calling model; self-host option (HF/Ollama) or watsonx hosted. Not a Grok replacement for frontier reasoning — good fallback / batch.
Pricing signal: Hosted e.g. granite-4-h-small ~USD 0.06 in / 0.25 out per 1M tokens; granite-3-8b-instruct ~$0.20/1M (indicative, check calculator).
URL: ibm.com/granite · Granite 4.1 docs

Model gateway (OpenAI-compat) must-try complement

What: GA path: OpenAI-compatible endpoints + OpenAI Python SDK for chat completions across providers on watsonx.
Donna relevance: Critical. Lets Donna’s proxy treat watsonx like another OpenAI base URL — multi-model routing without client rewrites.
Pricing signal: Billed as underlying model tokens / hosting; no separate “gateway tax” published beyond plan tiers.
URL: Model gateway docs

Prompt Lab · Tuning Studio · Guardrails maybe complement

What: Prompt engineering UI, LoRA/QLoRA fine-tune (Standard), Granite Guardian safety models (note some Guardian versions deprecated).
Donna relevance: Useful for offline eval of agent prompts; fine-tune only if you have domain data and Standard budget.
Pricing signal: LoRA e.g. ~$6.3/hr A100 / ~$14.85/hr H100 on Standard (indicative).
URL: watsonx.ai features

02 · Compute

Serverless · containers · VPS-like

IBM Cloud Code Engine must-try complement

What: Fully managed serverless: containers, source-to-deploy, functions, batch jobs; scale-to-zero. Serverless Fleets for GPU/VM parallel jobs (newer).
Donna relevance: Long-running agent jobs, webhooks, batch tools — not multi-tenant untrusted code sandboxes (still E2B) and not global edge Workers (still CF).
Pricing signal: Free tier ~100k vCPU-sec + 200k GB-sec + 100k HTTP req/mo. PAYG ~$0.000034/vCPU-s, ~$0.0000036/GB-s, ~$0.54/1M requests (third-party catalogs; confirm IBM calculator).
URL: code-engine · pricing docs

IBM Cloud Kubernetes Service skip ignore

What: Managed Kubernetes (+ OpenShift on IBM Cloud option). Cluster lifecycle, networking, GPU nodes.
Donna relevance: Opposite of Donna’s ops model. Only if an enterprise buyer mandates IBM K8s residency.
Pricing signal: Worker nodes billed as VSI/GPU + cluster management; occasional promo (e.g. 50% off containers 6mo — check current offers).
URL: IKS docs

Virtual Servers (VSI) / bare metal skip ignore

What: Classic and VPC virtual machines; bare metal for heavy dedicated hosts.
Donna relevance: No win vs specialized agent sandboxes. Legacy SoftLayer-era account UX still frustrates startups — avoid unless required.
Pricing signal: Hourly/monthly VSI SKUs; use IBM Cloud pricing calculator — do not invent unit rates.
URL: virtual-servers

Cloud Functions · Satellite skip ignore

What: Legacy Functions still exist in catalog narratives; Satellite extends IBM services to customer edge/on-prem.
Donna relevance: Prefer Code Engine over Functions. Satellite is enterprise hybrid — not a personal-agent startup need.
Pricing signal: Usage-based / enterprise contracts — check catalog if ever relevant.
URL: prefer Code Engine

03 · Data

Memory · vector · storage

watsonx.data (+ Milvus, OpenSearch) maybe ignore (now)

What: Open hybrid lakehouse: Presto/Spark engines, OpenRAG, integrated vector DB (Milvus path), OpenSearch hybrid search, governance.
Donna relevance: Enterprise RAG / agent knowledge at scale. Personal agent memory is better served by light Postgres/Redis/vector today.
Pricing signal: Flexible SaaS/software; free trial; see watsonx.data pricing page (workload-dependent).
URL: watsonx.data · vector DB announce

Databases for PostgreSQL / Redis maybe complement

What: Managed relational + cache on IBM Cloud; pgvector possible for light agent memory.
Donna relevance: Only if state lands near Code Engine jobs. CF-native path already covered.
Pricing signal: Plan-based (disk/RAM tiers); free lite plans exist on some IBM DB offerings — confirm catalog.
URL: IBM Cloud catalog

Object Storage · Event Streams maybe complement

What: S3-compatible COS; Kafka-compatible Event Streams for async agent events.
Donna relevance: Artifacts/logs for IBM-side jobs. Not required for CF+E2B core path.
Pricing signal: Storage + egress tiers; Event Streams partition/hour pricing — use calculator.
URL: Object Storage

04 · Agents

Orchestration & frameworks

watsonx Orchestrate maybe ignore (now)

What: Agentic control plane: multi-agent orchestration, Agent Builder, governed catalog, observability, hybrid deploy. ADK + MCP tool integration. Free trial / demo available.
Donna relevance: Enterprise HR/sales/IT automation GTM — different from multi-user personal Hermes. Partner Agent Connect could be distribution later.
Pricing signal: Usage/team scaled; free trial. See Orchestrate pricing page (no single public “$X/agent” to invent).
URL: Orchestrate · ADK

BeeAI Framework maybe complement

What: Open-source multi-agent framework (Linux Foundation governance). Declarative orchestration, MCP & A2A native, 10+ LLM providers incl. watsonx.ai. Integrates with Orchestrate as external agents.
Donna relevance: Interesting if rewriting agent runtime; optional experiment, not a forced migration from Hermes.
Pricing signal: Open source — free. Cost is your infra + model tokens.
URL: framework.beeai.dev · IBM overview

MCP (IBM work + ContextForge) maybe complement

What: MCP tools in Orchestrate ADK; ContextForge MCP gateway; docs MCP servers. Aligns with industry tool protocol.
Donna relevance: Composio already supplies tools. MCP is the open standard direction — useful interop, not an IBM lock-in reason.
Pricing signal: OSS / platform-included; no separate MCP SKU.
URL: MCP + Orchestrate tutorial

05 · Trust

Observability · security · IAM

watsonx.governance maybe complement

What: AI governance: evaluate/monitor models for accuracy, drift, bias, risk; newer agent monitoring & security metrics. GA enterprise platform (Gartner MQ AI governance leader narrative 2026).
Donna relevance: When selling into regulated enterprises wanting audit trails on agent decisions.
Pricing signal: Often paired with watsonx plans / separate governance entitlement — confirm on product pricing; trial via watsonx registration.
URL: watsonx.governance

IBM Cloud IAM · Secrets Manager maybe complement

What: Identity/access for IBM resources; dynamic secrets with single-tenant isolation options.
Donna relevance: Required if you run Code Engine/DBs on IBM. Not a substitute for multi-tenant app auth.
Pricing signal: IAM included with account; Secrets Manager has free/standard tiers — check catalog.
URL: Secrets Manager

Activity Tracker · Instana maybe complement

What: Audit events; full-stack APM with AI-agent traces (prompts, tokens, LangChain/LangGraph/CrewAI/watsonx — OpenLLMetry path).
Donna relevance: Instana is heavy vs lightweight OpenTelemetry + log drains Donna likely uses. Consider only at enterprise scale.
Pricing signal: Instana SaaS tiered; Activity Tracker event-based — calculator.
URL: Instana

06 · GTM

Startup · credits · free tiers

Startup with IBM must-try complement

What: Cloud credit program for early-stage software startups. Community + partner sources (2026): Builder ~$1,000/mo × 12; Premium ~$10,000/mo × 12 (VC/incubator path); up to ~$120k/yr.
Donna relevance: Highest-leverage IBM action: free credits for watsonx experiments without burn. Apply before becoming a paying IBM Cloud customer.
Pricing signal: Credits only — not cash. Eligibility commonly: <5 years, <$1M revenue, MVP, not prior credit recipient / not already paying customer (verify current form).
URL: GEP application · community thread

watsonx / IBM Cloud free tiers must-try complement

What: watsonx.ai free toolbox (~300k tokens, 20 CUH, 100 docs extraction). Code Engine monthly free allotment. Broader IBM Cloud Always Free catalog items.
Donna relevance: Prototype Granite + gateway with zero spend this week.
Pricing signal: As published on product free-tier pages — always re-check limits before planning load tests.
URL: IBM Cloud free · watsonx trial

Gotchas skip path

What: Legacy SoftLayer account UX, confusing classic vs VPC, sales-led enterprise pricing for Orchestrate/governance, regional model availability variance.
Donna relevance: Prefer modern VPC + Code Engine + watsonx SaaS. Watch APAC/Singapore region availability for latency-sensitive agents — verify per-model region matrix in docs before promising SLAs.
Pricing signal: N/A — process risk, not a product.
URL: cloud.ibm.com/docs

07 · Fit

Honest vs current Donna stack

E2B sandboxes — IBM has no peer multi-tenant untrusted code sandbox product. ignore / keep E2B
Cloudflare Workers gateway — Code Engine is regional serverless containers, not global edge. complement only
External LLMs (Grok/Codex) — watsonx gateway + Granite can sit beside them as alternate providers. complement
Composio connectors — Orchestrate catalog / MCP tools are parallel ecosystems; no need to rip Composio. interop later
Telegram-first UX — IBM products are studio/control-plane oriented; Donna keeps the product surface.
Replace anything now? — Nothing core. Credits + optional inference path only.

Bottom line: IBM is strongest as cheap/open models + credits + enterprise trust narrative, not as a full stack replacement for E2B + Cloudflare. Sell into IBM-shaped enterprises later; don’t rewrite Donna onto watsonx Orchestrate for consumers.

IBM vs current Donna stack

Layer-by-layer: keep, complement, or ignore.

Layer	Donna today	IBM offer	Verdict	Notes
Edge gateway	Cloudflare Workers	Code Engine apps	ignore	Keep CF for latency, DNS, Workers AI adjacency
Agent sandbox	E2B (Hermes)	Code Engine / VSI / K8s	ignore	No E2B-class multi-tenant firecracker UX
LLM proxy	External Grok / Codex	watsonx model gateway	complement	OpenAI-compat — add as provider
Foundation models	Frontier closed models	Granite 3/4 + 3rd-party on watsonx	complement	Cost control + open weights
Tools / connectors	Composio	Orchestrate catalog + MCP	complement	Parallel ecosystems; MCP is the bridge
Agent framework	Hermes runtime	BeeAI / Orchestrate ADK	ignore (now)	Optional experiment only
Memory / vector	App DB + embeddings	watsonx.data Milvus / Postgres	ignore (now)	Revisit for B2B knowledge bases
Secrets / IAM	CF secrets + app auth	Secrets Manager + IAM	complement	Only with IBM compute footprint
Observability	Logs / OTel-lite	Instana + watsonx.governance	complement	Enterprise sales motion
Credits / GTM	Bootstrap	Startup with IBM + free tiers	complement	Highest ROI first move
Full stack migrate	—	All-in watsonx	ignore	Wrong product shape for Telegram personal agents

Gaps & IBM weaknesses for this use case

Where IBM does not help Donna today — stated plainly.

No E2B-class sandbox

Nothing matches multi-tenant microVM sandboxes for untrusted agent code with pause/resume economics. Code Engine is trusted-workload serverless.

Not edge-native

No Cloudflare-style global Workers + KV/D1/R2 story. Latency-sensitive Telegram gateways stay on CF.

Enterprise gravity

Orchestrate, governance, and sales motions optimize for large IT buyers — slow for a 2–10 person startup shipping weekly.

Region / model matrix friction

Model availability and data residency vary by region. APAC latency needs explicit validation before promising agent SLAs.

Account UX debt

Classic vs VPC, SoftLayer-era confusion, and fragmented consoles cost founder time. Prefer watsonx SaaS + Code Engine only.

Frontier model lag

Granite is strong open/tool-call; still not a full substitute for top closed frontier models Donna already uses.

Recommended 30-day experiment: (1) Apply Startup with IBM credits. (2) Wire watsonx OpenAI-compatible gateway as a secondary provider in the LLM proxy. (3) Benchmark Granite 4.x tool-calling vs current default on real Donna tasks. (4) Deploy one non-critical batch job on Code Engine free tier. (5) Do not migrate sandboxes, Workers, or Composio.

watsonx.ai docs hub Granite docs Code Engine getting started Orchestrate overview BeeAI framework

Primary sources

Multi-source research: Exa neural search (api.exa.ai), IBM official product/docs pages, Grok web fetch, and X semantic search on watsonx / Granite / agents (2024–2026 chatter). Prefer primary docs over SEO blogs. Pricing is indicative — always re-check IBM calculators.

watsonx.ai product watsonx.ai pricing Model gateway (OpenAI-compat) IBM Granite Granite 4.1 models Code Engine Code Engine pricing watsonx.data watsonx Orchestrate BeeAI Framework watsonx.governance Secrets Manager Startup with IBM application IBM Cloud free tier Orchestrate ADK MCP + Orchestrate