# Agent Adapter

> A framework-independent Python SDK (3.11+). Write a tool once as a Python function and use it with any model API (OpenAI, Gemini, Anthropic), any agent framework (LangGraph, OpenAI Agents SDK, Pydantic AI, CrewAI, ...), MCP, A2A, AG-UI or HTTP. One executor adds validation, hooks, human approval, budgets, caching and tracing.

Import rules: everyday names come from `agent_adapter`; every other error from `agent_adapter.errors`; everything else from its subpackage (`agent_adapter.exec`, `.hooks`, `.approval`, ...); framework exports from `agent_adapter.integrations.<framework>`. Never import from `agent_adapter.core` or a `_private` module.

Full guide for AI coding agents (install, recipes, every public name): <https://agent-adapter.veloceai.in/llms-full.txt>

## Home

- [Home](https://agent-adapter.veloceai.in/): Agent Adapter is an open-source (Apache-2.0), framework-independent Python SDK (version **1.1.0**)

## Get started

- [Install](https://agent-adapter.veloceai.in/get-started/install/): Agent Adapter is on PyPI under the Apache-2.0 licence
- [Quickstart](https://agent-adapter.veloceai.in/get-started/quickstart/): Start with one function and one call
- [Wrap existing code](https://agent-adapter.veloceai.in/get-started/wrap-existing-code/): Use `registry.wrap()` for code you cannot or will not edit
- [First flow](https://agent-adapter.veloceai.in/get-started/first-flow/): Build a three-node flow that classifies a support ticket, then either looks up a policy with a tool or escalates
- [For AI coding agents](https://agent-adapter.veloceai.in/get-started/for-ai-agents/): One file teaches an AI coding agent (Claude Code, Cursor, Copilot) the SDK: what to install, which function does each job, the safety rules and the common...
- [Which feature do I need?](https://agent-adapter.veloceai.in/choose/): Answer the questions top to bottom
- [Features](https://agent-adapter.veloceai.in/features/): What Agent Adapter does, grouped by area, with a link to each guide
- [Troubleshooting](https://agent-adapter.veloceai.in/troubleshooting/): Find your symptom, read the cause, apply the fix

## How-to

- [Attach anywhere](https://agent-adapter.veloceai.in/how-to/attach-anywhere/): Write a tool once
- [Wrap a legacy class](https://agent-adapter.veloceai.in/how-to/wrap-legacy-code/): Turn an existing service class into tools without editing it: its methods, its blocking sync calls and its own exceptions
- [Parallel retrieval](https://agent-adapter.veloceai.in/how-to/parallel-retrieval/): Ask several retrievers at once, merge what they return, and stop early when one answer is enough
- [Human in the loop](https://agent-adapter.veloceai.in/how-to/human-in-the-loop/): Pause a risky call until a person reviews it, then resume it: approved, edited or denied
- [Run flows on Temporal or DBOS](https://agent-adapter.veloceai.in/how-to/durable-engines/): A `Flow` can run as a Temporal workflow or a DBOS workflow
- [Budgets](https://agent-adapter.veloceai.in/how-to/budgets/): Cap what a run or a conversation may spend, and how many calls it may make
- [Team budgets](https://agent-adapter.veloceai.in/how-to/team-budgets/): Cap what each user, each team, or everyone may spend, per UTC day or month, across every worker and host
- [Caching](https://agent-adapter.veloceai.in/how-to/caching/): Skip repeat work
- [Bring your own model](https://agent-adapter.veloceai.in/how-to/bring-your-own-model/): The adapter does not care which model runs
- [Model recipes](https://agent-adapter.veloceai.in/how-to/model-recipes/): Each recipe wraps the vendor's own SDK in a `CallableModel`
- [Send traces](https://agent-adapter.veloceai.in/how-to/send-traces/): `OTelSink` turns the executor's events into OpenTelemetry spans that follow the GenAI semantic conventions
- [Compress a long context](https://agent-adapter.veloceai.in/how-to/compress-long-context/): Your backend calls one model and the history keeps growing
- [Sync code and backends](https://agent-adapter.veloceai.in/how-to/sync-backends/): The adapter is async
- [Testing tools](https://agent-adapter.veloceai.in/how-to/testing-tools/): Test tools offline: no network, no API keys, no real waiting
- [See your runs](https://agent-adapter.veloceai.in/how-to/see-your-runs/): Record every run to a folder with one line, then open a local page that shows it: the flow graph, a timeline of parallel calls, costs, cache hits, retries...
- [Search and tag runs](https://agent-adapter.veloceai.in/how-to/search-your-runs/): The runs list of `agent-adapter ui` has a search box and a tag filter
- [Navigate the graph](https://agent-adapter.veloceai.in/how-to/navigate-the-graph/): The Agent Adapter run viewer draws the `Flow` a `FlowRunner` ran and paints each node's live status on it
- [Chat with your agent](https://agent-adapter.veloceai.in/how-to/chat-with-your-agent/): Chat with any agent in the local Agent Adapter page, and watch it work: streamed text, tool calls, approvals, the flow graph, a timeline, state and cost for...
- [Remember users across chats](https://agent-adapter.veloceai.in/how-to/chat-memory/): The chat page can remember what a user said in earlier chats
- [Chat flows by example](https://agent-adapter.veloceai.in/how-to/chat-flows/): `examples/chat_flows.py` is a scripted city help desk (no model, no API key): one agent picks a `Flow` per question, one per pattern
- [Wrap a framework for chat](https://agent-adapter.veloceai.in/how-to/chat-wrappers/): The Agent Adapter chat page talks to one function: `agent(message, session)`, returning text or an async iterator of text chunks (Chat with your agent)
- [Multi-agent](https://agent-adapter.veloceai.in/how-to/multi-agent/): Any agent can be a tool of another
- [Evaluate your agent](https://agent-adapter.veloceai.in/how-to/evaluate-your-agent/): Evals run a set of cases on your agent, grade each answer, and report
- [Evals from production](https://agent-adapter.veloceai.in/how-to/evals-from-production/): Turn recorded runs into eval cases, so a failure seen in production becomes a regression case that CI runs on every change
- [See your evals](https://agent-adapter.veloceai.in/how-to/see-your-evals/): Save each eval report to a folder, then open the **Evals** tab of the local viewer
- [Optimise your agent](https://agent-adapter.veloceai.in/how-to/optimise-your-agent/): The optimiser tunes the texts around any agent you can call: a plain function, or a LangGraph, OpenAI Agents or Pydantic AI agent
- [Cut model costs per step](https://agent-adapter.veloceai.in/how-to/cut-model-costs/): Not every step of an agent needs your best model
- [Tutorial: a research agent](https://agent-adapter.veloceai.in/tutorials/research-agent/): You build a small research agent end to end: tools from plain functions, retrieval from three sources at once, a model loop, a human approval before it...
- [Tutorial: the golden path](https://agent-adapter.veloceai.in/tutorials/golden-path/): A small support-desk service wired the way we recommend for production: a confirmed refund, approvals that survive a restart, a daily budget per user,...
- [Cookbook](https://agent-adapter.veloceai.in/cookbook/): Short recipes for tasks people ask about
- [Retry a flaky tool](https://agent-adapter.veloceai.in/cookbook/retry-flaky-tool/): A billing API drops connections now and then
- [Cap spend per user](https://agent-adapter.veloceai.in/cookbook/spend-per-user/): Each user may spend $0.25 a day on model calls
- [Redact arguments and results](https://agent-adapter.veloceai.in/cookbook/redact/): Keep API keys and emails out of your logs
- [Parallel retrieval with a deadline](https://agent-adapter.veloceai.in/cookbook/parallel-deadline/): Ask three retrievers at once
- [Approval over Slack or email](https://agent-adapter.veloceai.in/cookbook/approval-notify/): A wire transfer needs a person's yes
- [Cache model answers across workers](https://agent-adapter.veloceai.in/cookbook/cache-llm/): A FAQ bot gets the same question all day
- [Test tools with fakes](https://agent-adapter.veloceai.in/cookbook/test-with-fakes/): A tool calls a CRM
- [A registry per tenant](https://agent-adapter.veloceai.in/cookbook/tenant-registry/): Tenants on the basic plan get search only

## Examples

- [Gallery](https://agent-adapter.veloceai.in/examples/): Every example file under `examples/`, grouped by area
- [Basics](https://agent-adapter.veloceai.in/examples/basics/): Tools, the registry, schemas, the executor, models and guides
- [No-framework backends](https://agent-adapter.veloceai.in/examples/backends/): Web apps, workers, scripts, serverless, deploy
- [Agent frameworks](https://agent-adapter.veloceai.in/examples/frameworks/): Registry tools inside each agent framework
- [Context compression](https://agent-adapter.veloceai.in/examples/compression/): Shrink what each model call sends
- [Approval](https://agent-adapter.veloceai.in/examples/approval/): Human review, durable stores, resume once
- [Cost and cache](https://agent-adapter.veloceai.in/examples/cost-cache/): Ledgers, prices, budgets, caches, embeddings
- [Parallel fan-out and fan-in](https://agent-adapter.veloceai.in/examples/parallel/): Several calls at once, then merge
- [Memory and retrieval](https://agent-adapter.veloceai.in/examples/memory/): Per-user memory, and RAG with citations
- [Flows](https://agent-adapter.veloceai.in/examples/flows/): Graphs of nodes with joins, loops and pauses
- [Serving and protocols](https://agent-adapter.veloceai.in/examples/serving/): MCP, A2A, AG-UI, HTTP and OpenAPI
- [Evals and optimiser](https://agent-adapter.veloceai.in/examples/optimiser/): Grade an agent on eval cases; improve its harness
- [Cookbook](https://agent-adapter.veloceai.in/examples/cookbook/): Short recipes for common production needs

## Backends

- [FastAPI](https://agent-adapter.veloceai.in/backends/fastapi/): FastAPI runs on an event loop, so routes `await` the executor directly
- [Flask](https://agent-adapter.veloceai.in/backends/flask/): Flask views are sync
- [Django](https://agent-adapter.veloceai.in/backends/django/): Async views `await` the service
- [Celery](https://agent-adapter.veloceai.in/backends/celery/): Celery tasks are sync functions
- [Scripts and batch jobs](https://agent-adapter.veloceai.in/backends/scripts/): A CLI or a cron job owns its process and its loop
- [Serverless functions](https://agent-adapter.veloceai.in/backends/serverless/): A function host gives you a sync or async entry point and kills the container between invocations
- [Queues and consumers](https://agent-adapter.veloceai.in/backends/queues/): Every queue gives you a message body and a delivery id
- [Data pipelines](https://agent-adapter.veloceai.in/backends/pipelines/): Airflow, Prefect, Dagster, Ray and Spark run your code on workers, often in other processes
- [Other web frameworks](https://agent-adapter.veloceai.in/backends/other-asgi/): Any async framework can `await executor.run(...)` in its handlers, as the FastAPI guide does
- [Notebooks and apps](https://agent-adapter.veloceai.in/backends/notebooks/): Jupyter, Streamlit and Gradio already run an event loop, each in its own way
- [Deploy (Docker, Kubernetes)](https://agent-adapter.veloceai.in/backends/deploy/): Run the HTTP tool server (`create_app`) in a container

## Connect

- [Cross-language](https://agent-adapter.veloceai.in/cross-language/): Your tools stay Python functions in one `Registry`
- [MCP](https://agent-adapter.veloceai.in/cross-language/mcp/): MCP over **stateless Streamable HTTP** is the universal export
- [HTTP server](https://agent-adapter.veloceai.in/cross-language/http-server/): REST + SSE over one `Registry`, for what MCP cannot carry: approval decided in Python (with edited arguments), run events, cost per run, provider-shaped...
- [OpenAPI](https://agent-adapter.veloceai.in/cross-language/openapi/): The server publishes an **OpenAPI 3.1** document generated from the registry, so other languages can generate typed clients
- [TypeScript client](https://agent-adapter.veloceai.in/cross-language/typescript-client/): A zero-dependency client for the HTTP tool server
- [Connect anywhere](https://agent-adapter.veloceai.in/connect/): One page per stack
- [Hermes Agent](https://agent-adapter.veloceai.in/connect/hermes/): Hermes Agent (Nous Research) has a built-in MCP client
- [Letta](https://agent-adapter.veloceai.in/connect/letta/): Letta is a server platform
- [Mastra](https://agent-adapter.veloceai.in/connect/mastra/): Two ways in
- [Spring AI](https://agent-adapter.veloceai.in/connect/spring-ai/): Spring AI's MCP client Boot starter connects at start-up and turns each of our tools into a `ToolCallback`
- [MAF .NET](https://agent-adapter.veloceai.in/connect/maf-dotnet/): MAF .NET uses the official MCP C# SDK
- [Genkit](https://agent-adapter.veloceai.in/connect/genkit/): Genkit's `@genkit-ai/mcp` plugin runs an MCP host inside your Genkit app
- [ADK Go](https://agent-adapter.veloceai.in/connect/adk-go/): ADK Go's `mcptoolset` turns an MCP server into a `tool.Toolset` for an `llmagent`
- [LangGraph.js](https://agent-adapter.veloceai.in/connect/langgraph-js/): `@langchain/mcp-adapters` (2.0 or later) turns our MCP tools into LangChain.js tools
- [Vercel AI SDK](https://agent-adapter.veloceai.in/connect/vercel-ai-sdk/): AI SDK 7 (TypeScript)
- [OpenAI hosted MCP](https://agent-adapter.veloceai.in/connect/openai-hosted-mcp/): With the Responses API `mcp` tool, OpenAI's servers call our server for you
- [Azure AI Foundry](https://agent-adapter.veloceai.in/connect/azure-foundry/): A Foundry agent's `mcp` tool is run by the Foundry service: Azure calls our server, so it must be reachable from the internet, with auth
- [Bedrock AgentCore](https://agent-adapter.veloceai.in/connect/bedrock-agentcore/): AgentCore Gateway puts our server behind one MCP endpoint for your agents

## Integrations

- [Integrations](https://agent-adapter.veloceai.in/integrations/): Each integration is one module under `agent_adapter.integrations` and, where it needs a framework, one extra (install with `pip install...
- [Support matrix](https://agent-adapter.veloceai.in/support-matrix/): Every agent-framework integration and extra, its pinned range, and the nox session that tests it
- [Support matrix: backends](https://agent-adapter.veloceai.in/support-backends/): Every backend and runtime that needs no agent framework: how it attaches, its extra, its guide and the test that runs it
- [Compression by framework](https://agent-adapter.veloceai.in/integrations/compression-by-framework/): Every framework keeps its own message list
- [OpenAI](https://agent-adapter.veloceai.in/integrations/openai/): The renderers return plain dicts
- [Gemini](https://agent-adapter.veloceai.in/integrations/gemini/): Without google-genai, `genai_tool` and `run_function_calls` raise `MissingDependency` (also an `ImportError`)
- [Anthropic](https://agent-adapter.veloceai.in/integrations/anthropic/): The renderers return plain dicts
- [LangChain](https://agent-adapter.veloceai.in/integrations/langchain/): LangChain is an export target and an import source
- [OpenAPI](https://agent-adapter.veloceai.in/integrations/openapi/): `import_openapi` turns every operation of an OpenAPI 3.x document into a tool
- [MCP](https://agent-adapter.veloceai.in/integrations/mcp/): Serve registry tools to any MCP client, and import tools from any MCP server
- [MCP tasks and OAuth](https://agent-adapter.veloceai.in/integrations/mcp-tasks/): Three MCP 2026-07-28 features on top of MCP:
- [OpenAI Agents SDK](https://agent-adapter.veloceai.in/integrations/openai-agents/): Every registry tool becomes a `FunctionTool` that runs through the executor: validation, hooks, guards, cache, ledger and events apply inside the Agents SDK...
- [Pydantic AI](https://agent-adapter.veloceai.in/integrations/pydantic-ai/): Every registry tool becomes a Pydantic AI `Tool` (built with `Tool.from_schema` from the tool's neutral JSON schema) that runs through the executor
- [LlamaIndex](https://agent-adapter.veloceai.in/integrations/llamaindex/): Every registry tool becomes a LlamaIndex `FunctionTool` with `fn_schema=spec.args_model`
- [CrewAI](https://agent-adapter.veloceai.in/integrations/crewai/): Every registry tool becomes a CrewAI `BaseTool` with `args_schema=spec.args_model`
- [Microsoft Agent Framework](https://agent-adapter.veloceai.in/integrations/agent-framework/): The tool takes the neutral JSON schema as `input_model`
- [Agno](https://agent-adapter.veloceai.in/integrations/agno/): The function takes the neutral JSON schema as `parameters` (`skip_entrypoint_processing=True`)
- [Strands Agents](https://agent-adapter.veloceai.in/integrations/strands/): Results are Strands tool results: `status` `success` with a `text` (strings) or `json` block, or `status` `error`
- [Haystack](https://agent-adapter.veloceai.in/integrations/haystack/): `Agent.run_async` awaits the tool; `Agent.run` (sync) runs it through `run_sync`
- [Claude Agent SDK](https://agent-adapter.veloceai.in/integrations/claude-agent/): Claude sees each tool as `mcp__<server>__<tool>`, where `<server>` is the bridge's `name` (use it as the `mcp_servers` key)
- [Semantic Kernel](https://agent-adapter.veloceai.in/integrations/semantic-kernel/): Each parameter carries its property schema (`$ref`s inlined) as `schema_data`
- [DSPy](https://agent-adapter.veloceai.in/integrations/dspy/): DSPy 3.4+ pulls in litellm, openai and boto3 (about 300 MB), so the extra is not part of `all` and its contract tests run in their own `nox -s...
- [smolagents](https://agent-adapter.veloceai.in/integrations/smolagents/): smolagents tools are synchronous
- [AutoGen](https://agent-adapter.veloceai.in/integrations/autogen/): For AutoGen 0.7 (`autogen-core`, `autogen-agentchat`)
- [LangGraph](https://agent-adapter.veloceai.in/integrations/langgraph/): Registry tools run inside LangGraph through an `Executor`, so hooks, approval, timeouts, retries, the cache and the ledger apply (`executor=None` uses the...
- [Google ADK](https://agent-adapter.veloceai.in/integrations/adk/): Declarations use the Gemini schema family (`to_gemini_parameters`)
- [CAMEL-AI](https://agent-adapter.veloceai.in/integrations/camel/): The schema is the OpenAI transform of the neutral schema
- [AG2](https://agent-adapter.veloceai.in/integrations/ag2/): AG2 1.x is the `ag2` package (`from ag2 import Agent`)
- [A2A](https://agent-adapter.veloceai.in/integrations/a2a/): Serve registry tools as an A2A agent, and import a remote A2A agent's skills as tools
- [AG-UI](https://agent-adapter.veloceai.in/integrations/ag-ui/): Frameworks that ship their own AG-UI adapter (Pydantic AI, LangGraph, ADK, MAF, CrewAI) already carry our tools inside them

## Concepts

- [Architecture](https://agent-adapter.veloceai.in/architecture/): The core is small and depends only on pydantic
- [Tool lifecycle](https://agent-adapter.veloceai.in/architecture/tool-lifecycle/): A tool goes through two phases
- [Schema pipeline](https://agent-adapter.veloceai.in/architecture/schema-pipeline/): Each tool has one JSON schema, built by pydantic from its args model
- [Executor and flows](https://agent-adapter.veloceai.in/architecture/executor-and-flows/): Every tool call runs through one `Executor` pipeline
- [HITL, cost and cache](https://agent-adapter.veloceai.in/architecture/hitl-cost-cache/): How approval, the cost ledger, the caches and the model ports fit together around one agent step
- [Tools](https://agent-adapter.veloceai.in/concepts/tools/): A **tool** is a callable plus what an LLM needs to call it: a name, a description and an argument schema
- [Registry](https://agent-adapter.veloceai.in/concepts/registry/): A `Registry` holds tools by name, in registration order
- [Schemas](https://agent-adapter.veloceai.in/concepts/schemas/): Each tool has one JSON schema, built by pydantic
- [Strict mode and nulls](https://agent-adapter.veloceai.in/concepts/strict-mode/): OpenAI strict mode (the default for `chat_tools` and `responses_tools`) has two rules:
- [Context, events and ports](https://agent-adapter.veloceai.in/concepts/context/): A tool that declares a `ToolContext` parameter gets one at call time
- [Runtime](https://agent-adapter.veloceai.in/concepts/runtime/): `agent_adapter.runtime` holds the pieces every async API leans on: a sync facade, values kept per event loop, the current context and ids
- [Executor](https://agent-adapter.veloceai.in/concepts/executor/): `Executor` (`from agent_adapter import Executor`) runs tools
- [Executor policies](https://agent-adapter.veloceai.in/concepts/executor-policies/): What the Executor enforces on every call: approval, run guards, retries and timeouts, the tool cache, the ledger and events
- [Flows](https://agent-adapter.veloceai.in/concepts/flows/): A flow is a graph of nodes over one pydantic state
- [Flow pause and resume](https://agent-adapter.veloceai.in/concepts/flow-resume/): A flow pauses at a `pause` gate or when a tool inside a node needs an approval that is not decided yet
- [Sub-flows and node cache](https://agent-adapter.veloceai.in/concepts/flow-subflows/): A `SubFlow` node runs another `Flow` as one node of its parent
- [Durable runs](https://agent-adapter.veloceai.in/concepts/durable-runs/): A durable run survives its worker
- [Hooks](https://agent-adapter.veloceai.in/concepts/hooks/): A hook is middleware around every tool call, model call and flow node
- [Guardrails](https://agent-adapter.veloceai.in/concepts/guardrails/): A guard is a hook that looks at what goes into and out of your agent: tool arguments, tool results, model requests and model replies
- [Policy file](https://agent-adapter.veloceai.in/concepts/policy/): A policy is one file that says, per tool, per team or user and per environment, which tools may run, which need a person's approval, which guards check...
- [Audit log](https://agent-adapter.veloceai.in/concepts/audit/): `agent_adapter.audit` keeps a signed, tamper-evident record of the security-relevant things your agents did: approvals asked, decided and used, denials,...
- [Audit log in production](https://agent-adapter.veloceai.in/concepts/audit-operations/): How the audit log behaves under load and failure, what strict mode stops, and what the log cannot prove on its own
- [Approval](https://agent-adapter.veloceai.in/concepts/approval/): A tool with `confirm=` set asks an approval port before it runs
- [Resume a parked call](https://agent-adapter.veloceai.in/concepts/approval-resume/): A `confirm` tool behind `DeferredApproval` pauses with `ApprovalPending`
- [Approval batches](https://agent-adapter.veloceai.in/concepts/approval-batches/): A framework that pauses a run for several tool calls at once (the Agents SDK, Pydantic AI) asks the approval port about the whole batch
- [Approval stores](https://agent-adapter.veloceai.in/concepts/approval-stores/): A `DeferredApproval` parks a call in an `ApprovalStore` until a reviewer decides
- [Database approval stores](https://agent-adapter.veloceai.in/concepts/approval-stores-db/): Two more approval stores keep parked calls in a database you already run: **Postgres** through SQLAlchemy 2 (async), and **DynamoDB** for AWS serverless...
- [Approval memory](https://agent-adapter.veloceai.in/concepts/approval-memory/): Reviewers decide the same kinds of calls again and again
- [Cost](https://agent-adapter.veloceai.in/concepts/cost/): `agent_adapter.cost` turns token usage into dollars
- [Tracing](https://agent-adapter.veloceai.in/concepts/tracing/): Every step emits an `Event` (catalogue in Context, events and ports) to one `EventSink`
- [Cache](https://agent-adapter.veloceai.in/concepts/cache/): `agent_adapter.cache` saves money and time at four layers
- [Embeddings](https://agent-adapter.veloceai.in/concepts/embeddings/): `agent_adapter.embeddings` turns texts into vectors with **your** embedding model
- [Memory](https://agent-adapter.veloceai.in/concepts/memory/): `agent_adapter.memory` lets an agent remember facts across turns and chats
- [Retrieval (RAG)](https://agent-adapter.veloceai.in/concepts/retrieval/): `agent_adapter.retrieval` lets a model search **your** documents and answer with citations
- [Sandboxed code](https://agent-adapter.veloceai.in/concepts/sandbox/): `agent_adapter.sandbox` lets a model write code and run it somewhere it can do no harm
- [Model ports](https://agent-adapter.veloceai.in/concepts/models/): A model port turns one provider-neutral `LLMRequest` into one `LLMResponse`
- [Streaming](https://agent-adapter.veloceai.in/concepts/streaming/): A model port's `stream(request)` yields events while the reply arrives
- [Instrumented models and tool loop](https://agent-adapter.veloceai.in/concepts/tool-loop/): `InstrumentedModel` wraps any `LLMPort` with the executor's hooks, the LLM response cache, one ledger entry per call and `llm.*` events
- [Context compression](https://agent-adapter.veloceai.in/concepts/context-compression/): Long conversations hit the model's context window and cost more on every call
- [Errors](https://agent-adapter.veloceai.in/concepts/errors/): Symptoms and fixes: Troubleshooting
- [Optimiser](https://agent-adapter.veloceai.in/optimiser/): The optimiser improves an agent by editing its **harness** (tool descriptions and hints, hook settings, system prompt sections, flow node prompts), never...
- [Glossary](https://agent-adapter.veloceai.in/concepts/glossary/): One line per term, in alphabetical order

## Guides

- [Security](https://agent-adapter.veloceai.in/guides/security/): What can go wrong when a model drives your tools, what the adapter stops on its own, and what your app must still do
- [Threat model](https://agent-adapter.veloceai.in/guides/threat-model/): What an agent built on the adapter has to protect, where the trust boundaries are, what the adapter enforces on its own, and what stays your job
- [Production checklist](https://agent-adapter.veloceai.in/guides/production-checklist/): Go through this list before an agent built on the adapter takes real traffic
- [Performance](https://agent-adapter.veloceai.in/guides/performance/): Where time goes, and the knobs that change it
- [Observability](https://agent-adapter.veloceai.in/guides/observability/): See what a run did, what it cost, and why it failed, without leaking its data
- [Versioning](https://agent-adapter.veloceai.in/guides/versioning/): Agent Adapter follows Semantic Versioning 2.0
- [Migration guides](https://agent-adapter.veloceai.in/migration/): Move existing tool code onto the adapter one step at a time
- [Raw tool calling](https://agent-adapter.veloceai.in/migration/raw-tool-calling/): You wrote the JSON schema by hand, kept a `name -> function` dict, and called `fn(**json.loads(arguments))`
- [LangChain tools](https://agent-adapter.veloceai.in/migration/langchain-tools/): Your tools are LangChain `@tool` functions
- [Hand-rolled approval](https://agent-adapter.veloceai.in/migration/hand-rolled-approval/): You check "does this need approval?" inside the tool, park the arguments in a dict or a table, and run the tool once someone flips a flag
- [The migrate command](https://agent-adapter.veloceai.in/migration/migrate-command/): `migrate` reads your Python files and lists every framework tool definition it finds
- [From LangGraph](https://agent-adapter.veloceai.in/tutorials/migrate-langgraph/): You keep your graph, your state and your checkpointer

## Reference

- [API reference](https://agent-adapter.veloceai.in/reference/): Generated from the docstrings
- [Top level](https://agent-adapter.veloceai.in/reference/top-level/): The everyday names, imported from `agent_adapter`
- [Errors](https://agent-adapter.veloceai.in/reference/errors/): Every error Agent Adapter raises, in one table
- [Executor](https://agent-adapter.veloceai.in/reference/exec/): Running tool calls: retries, plans and the tool loop
- [Hooks](https://agent-adapter.veloceai.in/reference/hooks/): Hooks around tool, model and node calls, the built-in hooks and content guards
- [Approval](https://agent-adapter.veloceai.in/reference/approval/): Confirm rules, approval ports and stores
- [Flows](https://agent-adapter.veloceai.in/reference/flow/): Flow graphs, nodes and the runner
- [Cost](https://agent-adapter.veloceai.in/reference/cost/): Ledger sinks, summaries and resolved prices
- [Cache](https://agent-adapter.veloceai.in/reference/cache/): Tool-result and model-response caches, keys and provider prompt caches
- [Embeddings](https://agent-adapter.veloceai.in/reference/embeddings/): The embedding port, a callable embedder and its cache
- [Memory](https://agent-adapter.veloceai.in/reference/memory/): The memory port, SQLite and callable backends, Mem0/Zep/Graphiti adapters and the memory tools
- [Retrieval](https://agent-adapter.veloceai.in/reference/retrieval/): The retriever port, documents with sources, callable, embedding, LlamaIndex and LangChain retrievers, and `retriever_tool`
- [Sandbox](https://agent-adapter.veloceai.in/reference/sandbox/): The sandbox port, Docker, callable and development-only local backends, and `code_tool`
- [Policy](https://agent-adapter.veloceai.in/reference/policy/): One policy file for access, approvals, guards, budgets and limits per tool, team and environment
- [Audit log](https://agent-adapter.veloceai.in/reference/audit/): A signed, tamper-evident log of approvals, guard blocks and redactions, budget stops, memory deletes and exports
- [Context compression](https://agent-adapter.veloceai.in/reference/context/): Compression policies, steps, scorers, `CompressingModel` and the message-format converters
- [Runtime](https://agent-adapter.veloceai.in/reference/runtime/): The current tool context, ids and loop-local state
- [Tracing](https://agent-adapter.veloceai.in/reference/tracing/): Events, event sinks and OpenTelemetry spans
- [Models](https://agent-adapter.veloceai.in/reference/models/): Model ports, provider models and usage helpers
- [HTTP server](https://agent-adapter.veloceai.in/reference/server/): The tool service, its ASGI and WSGI apps, auth and problems
- [Handlers](https://agent-adapter.veloceai.in/reference/handlers/): Run tool calls from queue messages, serverless events and pipelines
- [Chat contract](https://agent-adapter.veloceai.in/reference/chat/): The types a chat agent receives for one turn and what it may return, shared by the chat page and evals
- [Run viewer](https://agent-adapter.veloceai.in/reference/viewer/): Record runs to a folder and browse them on 127.0.0.1 (`agent-adapter ui`), or chat with an agent there
- [Integrations: model APIs and protocols](https://agent-adapter.veloceai.in/reference/integrations-providers/): Export tools to model APIs and agent protocols (MCP, A2A, AG-UI); import tools from LangChain, OpenAPI and MCP
- [Integrations: frameworks](https://agent-adapter.veloceai.in/reference/integrations-frameworks/): Export tools, approval and flows to agent frameworks
- [Integrations: durable engines](https://agent-adapter.veloceai.in/reference/integrations-durable/): Flows on Temporal and DBOS: Run flows on Temporal or DBOS
- [Integrations: protocol](https://agent-adapter.veloceai.in/reference/integrations-base/): The integration protocol and entry-point discovery, for authors of new integrations (Writing an integration)
- [Evals](https://agent-adapter.veloceai.in/reference/evals/): Run eval cases on any agent, grade them and report (Evaluate your agent)
- [Optimiser](https://agent-adapter.veloceai.in/reference/optimiser/): Improve tool descriptions and prompts against eval cases (Optimiser)
- [CLI](https://agent-adapter.veloceai.in/reference/cli/): The `agent-adapter` command has subcommands
- [Changelog](https://agent-adapter.veloceai.in/changelog/): All notable changes to this project are documented here

## Contributing

- [Writing an integration](https://agent-adapter.veloceai.in/contributing/writing-an-integration/): An integration turns registry tools into one framework's native tools
- [Performance regression tests](https://agent-adapter.veloceai.in/contributing/performance-tests/): A test that guards against exponential or quadratic work must not time the work: seconds change with machine load, so a `< 2 s` assertion flakes
- [Cutting a release](https://agent-adapter.veloceai.in/contributing/release/): How to ship a new version of Agent Adapter to PyPI
- [API stability policy](https://agent-adapter.veloceai.in/contributing/stability/): The rules every change to Agent Adapter follows in 1.x: what callers may rely on, what may change, and how something is retired
- [Roadmap](https://agent-adapter.veloceai.in/roadmap/): What we plan to add or improve after **1.0**