Niches · Agent frameworks & SDKs
Build an agent that calls the app's own functions and APIs
One company's docs show a full fit.
What counts as done
Functions become tools with schemas generated from code, the agent calls them and answers, the run stops on its own at a documented turn, tool-call or spend cap with a typed reason, and the same code runs against a second provider or a local endpoint. A test model or mock lets the wiring be tested with no key.
Measured, for our own tests: tool-call success on a fixed set of 20 tasks and the number of model calls per task. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: build an agent that can call our API
, make a chatbot that looks things up in our database
, turn these functions into tools for an agent
, add an assistant to my app
Does the job 1
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| AI SDK AA Vercel Inc. | ToolLoopAgent runs Zod-typed tools, stops after 20 steps by default or on a stopWhen condition, reports finishReason per step, runs on many providers including OpenAI-compatible local endpoints, and ai/test ships mock models for keyless tests. Measured success is the suite's. source |
Does the job, with a limit 16
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| AgentOS A Framers Lab, Inc. | Tools run in a loop on 13 providers including Ollama, and a budget refuses calls past maxCostUSD with CostCapExceededError. No turn or tool-call cap and no test or mock model found in the docs read. source |
| Agno BB Agno Inc. | Functions become tools and tool_call_limit caps calls per loop, but the default is None and an open issue says the cap blocks tools while the model keeps looping. No typed stop reason or test model found. source |
| BoxLang AI B Ortus Solutions, Corp. | Tools come from annotated functions and MaxToolCallsMiddleware enforces a per-run cap, per the docs index. I found no typed stop reason, and didn't confirm the listing's mock provider; BoxLang runs only on the JVM. source |
| Claude Agent SDK BB Anthropic, PBC | Custom tools run through an in-process MCP server; maxTurns and maxBudgetUsd end the run with a result subtype. Claude only (API, Bedrock, Google Cloud, Foundry), so no local OpenAI-compatible endpoint, and no mock model is documented. source |
| CrewAI C CrewAI, Inc. | Tools attach per agent or task and models route through many providers including local ones; I found no documented typed stop reason when an iteration cap is hit, and no test model. source |
| Cursor Agent SDK B Anysphere, Inc. | Tools take Zod input schemas, an effect class for dry runs and retries for idempotent calls. Models are Cursor's catalogue only, with no local endpoint, and the config reference shows a concurrency cap but no turn or spend cap. source |
| Docker Agent BB Docker, Inc. | Tools come from toolsets, OpenAPI and scripts rather than functions in host code, and budgets stop a run naming the limit that tripped. Every limit is unset by default, so a run is unbudgeted unless the config declares one. source |
| Agent Development Kit (ADK) BB Google LLC | Functions become tools with schemas from code, and LiteLLM and Ollama pages cover other providers. An account saw every Ollama model loop on tool calls without answering (adk-python #81); no turn cap with a typed stop reason was found in the pages read. source |
| LangGraph BB LangChain, Inc. | create_agent builds a tool loop on any LangChain chat model including local ones. An account found the graph DSL too much for simple agents (HN 48661127); I didn't confirm a typed stop reason for the recursion limit this run. source |
| LlamaIndex Framework BB LlamaIndex, Inc. | FunctionAgent turns Python functions into tools on any provider with tool calling, ReActAgent covers others. A typed stop on a turn or tool-call cap and a mock LLM weren't found in the pages read. source |
| Mastra A Kepler Software, Inc. | createTool takes Zod schemas and the agent loop runs on any AI SDK provider including local ones. I didn't confirm a typed stop reason on a step cap or a mock model this run; an account found the basics work for a local coding agent (HN 48530462). source |
| Microsoft Agent Framework A Microsoft Corporation | @tool functions become tools on several chat clients, and termination middleware can end runs. Tools run without approval unless set, and the docs read didn't show a typed stop reason on a turn cap or a mock client. source |
| OpenAI Agents SDK AA OpenAI OpCo, LLC | function_tool builds schemas from code, max_turns raises MaxTurnsExceeded or routes to an error_handlers entry, and an account ran local models through LM Studio (HN 48326951). The testing page doesn't name a fake model for keyless wiring tests. source |
| Pydantic AI AA Pydantic Services Inc. | Tools come from typed functions on many providers, and TestModel or FunctionModel swap in via Agent.override for keyless tests. An account saw the model search more than 10 times a run (HN 45798996); the run caps and their stop type weren't re-read. source |
| smolagents B Hugging Face, Inc. | ToolCallingAgent emits JSON tool calls on any model including Ollama through LiteLLM, and final_answer_checks control termination. An account found tool calling beat code actions on current models (HN 46675615); no mock model or typed stop reason was found in the pages read. source |
| Strands Agents A Amazon Web Services, Inc. | Model-driven loop with decorated tools and many providers. An account's NBA agent made unasked calls, cost about 10 times the estimate and hit a Lambda timeout (AWS builder article); a turn cap with a typed stop reason wasn't found in the pages read. source |
Doesn't do this job 3
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| guilds.run C Cigale AI | Tools come from native tools and catalog connectors configured in the app, not from functions in the developer's own code; runs do stop on max steps, max tool calls and max input tokens. source |
| NitroStack C Nitrostack Inc. | It defines tools with input schemas for MCP clients to call but runs no model or agent loop itself. source |
| TradingAgents B Tauric Research | The agents call its own market-data tools; the README has no way to turn an app's functions into tools for a general agent. source |
What stops agents finishing
- a tool-call cap that blocks tools but leaves the loop running
- local models looping on tool calls with no stop
- a tool with no parameters sent in a shape strict servers reject
- breaking changes in minor releases of a 0.x package
Around it
Before: inference.tool-loop. After: agent-observability.instrument-app, frameworks.multi-agent-handoff.
Where this comes from
Weight 13 of 100 in Agent frameworks & SDKs: 8 of 61 accounts have this as their main job (share 13.1, rounded by largest remainder to sum 100)
The same as data: /api/v1/use-cases/frameworks.tool-calling-agent.json, or the MCP tool get_use_case with id frameworks.tool-calling-agent. Other jobs in this category: /api/v1/niches?category=frameworks.
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.