Starter stacks · model, framework, tools and data
Starter stacks
Starting points for common agents, one listing per job, picked from the rankings. A stack is only as good as its weakest part, so that's the grade we show for it. Swap anything you like. The point is that every piece has been checked.
Research agent
For an agent that answers questions from the web and shows its sources
A strong model, a typed framework that validates the answer, a search API built for agents, a reader that turns pages into Markdown, and an encyclopaedia with citations for the background.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Model | Claude APIAnthropic | BB | Best task-suite score in the run for multi-step research. Cache the system prompt and tool list. |
| Framework | Pydantic AIPydantic | A | Typed outputs mean the answer arrives with its sources as fields, not prose. No telemetry by default. |
| Search | Exa API + MCPExa | B | Search and page contents in one API, anonymous on the hosted MCP and payable per call. |
| Reader | Jina ReaderJina AI (Elastic) | E | Any URL to Markdown with no key, about a twentieth of the tokens of raw HTML. |
| Background | Wikimedia REST APIWikimedia Foundation | B | Free, keyless, and every article carries its own citations. |
Watch for
- Exa's deep modes cost two to three times the fast mode. Pick the mode per question
- Send Wikimedia a User-Agent with a contact address or the limits bite
- Jina bills in tokens with no dollar price on the page
Swaps
- Tavily for Exa if you want credits rather than per-call prices
- LangGraph for Pydantic AI when a run lasts hours and must survive restarts
- GPT-6 Sol for Claude if you're already on OpenAI
Through letme (picks today, calling later): letme.dev/anthropic-api · letme.dev/pydantic-ai · letme.dev/exa-mcp · letme.dev/jina-reader · letme.dev/wikimedia
Coding agent
For an agent that works in a repository, reads current docs, checks its work in a browser and reads production errors
Claude Code's own loop as a library, with the official GitHub server, library docs in the context window, a browser for checking the result and the error tracker for what broke.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Model | Claude APIAnthropic | BB | The model the Agent SDK is built for. |
| Framework | Claude Agent SDKAnthropic | BB | File, shell and search tools with permission modes, out of the box. |
| Repository | GitHub MCP ServerGitHub | BB | Official, hosted, with toolsets to keep the schema small. |
| Library docs | Context7Upstash | BB | Current docs for the library version in use, not the one the model remembers. |
| Browser check | Playwright MCPMicrosoft | B | Accessibility-tree snapshots cost far fewer tokens than screenshots. |
| Production errors | Sentry MCPSentry | BB | Reads the stack trace the change was meant to fix. |
Watch for
- Set
DISABLE_TELEMETRY=1for the Agent SDK - Context7's resolver description is long. Load it only when the agent needs docs
- Scope the GitHub server to one repository and a read-only toolset until you trust it
Swaps
- OpenAI Agents SDK with GPT-6 Sol if you need other models
- Chrome DevTools MCP for Playwright when the job is debugging performance
Through letme (picks today, calling later): letme.dev/anthropic-api · letme.dev/claude-agent-sdk · letme.dev/github-mcp-server · letme.dev/context7 · letme.dev/playwright-mcp · letme.dev/sentry-mcp
Operations and support agent
For an agent inside a company's own tools, working through customer records, tickets, chat and incidents with each user's own permissions
Everything here is OAuth inside the organisation's existing permissions, so the agent can do what the person could and nothing more.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Model | OpenAI APIOpenAI | A | Remote MCP in the Responses API with approval policies, and project-scoped keys. |
| Framework | OpenAI Agents SDKOpenAI | AA | Guardrails and human approval are first-class, and hand-offs suit triage. |
| CRM | HubSpot API + MCPHubSpot | BB | Official, hosted, OAuth, and the records most support work starts from. |
| Tickets and docs | Atlassian Rovo MCP ServerAtlassian | C | Jira and Confluence through one hosted server. |
| Chat | Slack MCP Server (official)Slack (Salesforce) | C | Read threads and post updates as the user. |
| Incidents | PagerDuty MCP ServerPagerDuty | D | On-call and incident state from the source of truth. |
Watch for
- Turn off tracing to OpenAI or add your own trace processor before customer data flows
- Atlassian switches remaining v1 connections to v2 tools on 2027-03-01
- PagerDuty's self-hosted server is archived. Use the hosted one
Swaps
- Claude API and Claude Agent SDK for OpenAI if your harness is already Claude
- Linear for Jira if that's where the work lives
Through letme (picks today, calling later): letme.dev/openai-api · letme.dev/openai-agents-sdk · letme.dev/hubspot-mcp · letme.dev/atlassian-rovo-mcp · letme.dev/slack-mcp · letme.dev/pagerduty-mcp
Market data agent
For an agent that answers questions about prices, companies and markets with data it can cite
Consolidated-tape US market data, digital-asset data from a regulated benchmark administrator, the UK company register, and a durable framework for jobs that run all day.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Model | OpenAI APIOpenAI | A | GPT-6 Luna is cheap enough to run on every tick, Sol for the analysis. |
| Framework | LangGraphLangChain | BB | Checkpoints, so a long-running monitor resumes after a restart. |
| Equities and options | Massive (formerly Polygon.io)Massive | BB | SIP and FINRA data with the sources named, and x402 for calls without an account. |
| Digital assets | CoinDesk Data APICoinDesk Data (formerly CCData / CryptoCompare) | D | Highest data-quality score in the directory for digital assets. Disclosure: built by Anchor Terminal's founder, see the listing. |
| Company facts | Companies House APICompanies House | D | The statutory UK register, free. |
Watch for
- Individual Massive plans are non-commercial. Check yours before an agent republishes a price
- CoinDesk has no free tier since 2026-05-21
- Companies House allows 600 requests per 5 minutes per key
Swaps
- CoinGecko's x402 API for CoinDesk when the agent has to pay per call with no account
- Pydantic AI for LangGraph if the jobs are short
Through letme (picks today, calling later): letme.dev/openai-api · letme.dev/langgraph · letme.dev/massive · letme.dev/coindesk-data-api · letme.dev/companies-house
Low cost, high volume
For an agent that makes thousands of small calls a day and has to stay cheap
Fast open-weight models for the many small calls, a router for the few hard ones, a framework that validates cheap models' output, and the cheapest search and reading in the directory.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Fast model | GroqCloudGroq | BB | GPT-OSS 20B at $0.075/$0.30 and about 1,000 tokens a second. |
| Hard calls | OpenRouterOpenRouter | B | Route the few hard calls to a frontier model with a max_price ceiling. |
| Framework | Pydantic AIPydantic | A | Output validation retries fix cheap models' malformed answers. |
| Search | Brave Search API + MCPBrave | B | $5 per 1,000 searches, published. |
| Reader | Jina ReaderJina AI (Elastic) | E | No key for 20 pages a minute. |
Watch for
- Groq retired models four times this quarter. Check
/modelsat start-up - OpenRouter's fee is 5.5% on card top-ups
- Brave needs a card even for the free tier
Swaps
- Gemini 3.8 Flash for Groq until its price doubles on 2027-01-01
- Mistral Small 4 for the fast model if the data must stay in the EU
Through letme (picks today, calling later): letme.dev/groq · letme.dev/openrouter · letme.dev/pydantic-ai · letme.dev/brave-search-mcp · letme.dev/jina-reader
European vendors
For teams that want their agent's vendors established in Europe
A French model provider with EU endpoints, a framework that sends nothing home, and data and scraping from companies in Germany and the Czech Republic. Established in Europe isn't the same as processing only in Europe, so read each listing's data details.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Model | Mistral AI APIMistral AI | BB | EU endpoint for 10% extra, and open weights you can run yourself later. |
| Framework | Pydantic AIPydantic | A | No telemetry by default, runs wherever you do. |
| Reader | Jina ReaderJina AI (Elastic) | E | Jina AI GmbH, Berlin, now part of Elastic. |
| Scraping | Apify MCP ServerApify | A | Apify Technologies s.r.o., Prague. |
| Weather | Open-MeteoOpen-Meteo | B | OpenMeteo GmbH, with European weather models among its sources. |
Watch for
- Mistral's Labs and preview models train on your data from 2026-09-25
- Jina's privacy statement is Elastic's
- Apify telemetry is on by default
Swaps
- Companies House for UK company data. The UK isn't in the EU, but it's often the question
Through letme (picks today, calling later): letme.dev/mistral-api · letme.dev/pydantic-ai · letme.dev/jina-reader · letme.dev/apify-mcp · letme.dev/open-meteo
Pays its own way
For an agent with its own small wallet that buys what it needs per call, with no accounts
Every paid call here settles with x402 from a capped wallet. It's the smallest stack on the page and the only one that needs no human signup at all.
| Job | Listing | Grade | Why this one |
|---|---|---|---|
| Protocol | x402x402 Foundation (Linux Foundation) | A | No account and no protocol fee. |
| Wallet | Coinbase Developer Platform (Agentic Wallet, AgentKit, CDP MCP)Coinbase Developer Platform | BB | Spend-capped wallet and payer tooling from the protocol's creator. |
| Model | BlockRun.AIBlockRun, Inc. | BB | OpenAI-compatible inference paid per request in USDC. |
| Search | Exa API + MCPExa | B | $0.007 a search over x402. |
| Market data | Massive (formerly Polygon.io)Massive | BB | x402 at agent.massive.com since September. |
Watch for
- Cap the wallet. $5 is plenty to start
- BlockRun's domain is under a year old and its people aren't named
- Check the amount and payee in every 402 before signing
Swaps
- CoinGecko's x402 API for Massive for digital-asset prices
- L402 sellers if the wallet holds bitcoin
Through letme (picks today, calling later): letme.dev/x402 · letme.dev/coinbase-cdp-agentkit · letme.dev/blockrun-ai · letme.dev/exa-mcp · letme.dev/massive
Stacks are editorial and never sponsored. /api/v1/stacks.json has them as data.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.