分享您的中间件
中间件支持上下文工程、 harness 自定义和运行时安全控制。它是 LangChain 中一个有用的扩展点，我们很乐意展示社区使用它所构建的内容：Add an official integration
按照贡献指南构建和发布中间件包。
Share a community middleware
向文档仓库提交 PR，将您的中间件添加到所有集成表格中。
精选集成
| 提供商 | 可用中间件 | 来源 | 下载 |
|---|---|---|---|
OpenAI middleware | 内容审核、模型路由和工具风险门控 | langchain-ai/langchain |  |
Anthropic middleware | 提示词缓存、bash 工具、文本编辑器、记忆和文件搜索 | langchain-ai/langchain |  |
AWS middleware | 提示词缓存和 AgentCore 支付 | langchain-ai/langchain-aws, aws/bedrock-agentcore-sdk-python |  |
Microsoft Foundry middleware | 文本审核、图像审核、提示词盾牌、受保护材料和对齐性 | langchain-ai/langchain-azure |  |
NVIDIA middleware | 模型路由和 Nemotron 3 Ultra 执行优化 | langchain-ai/langchain-nvidia, langchain-ai/deepagents | 不适用 |
所有中间件
| Provider | Middleware available | Source | Downloads |
|---|---|---|---|
OpenAI middleware | Content moderation, model routing, and tool-risk gating | langchain-ai/langchain | |
Anthropic middleware | Prompt caching, bash tool, text editor, memory, and file search | langchain-ai/langchain | |
AWS middleware | Prompt caching and AgentCore Payments | langchain-ai/langchain-aws, aws/bedrock-agentcore-sdk-python | |
Microsoft Foundry middleware | Text moderation, image moderation, prompt shield, protected material, and groundedness | langchain-ai/langchain-azure | |
CopilotKit | CopilotKit middleware and FastAPI bridge for Deep Agents, create_agent graphs, AG-UI, and the React and runtime clients | CopilotKit/CopilotKit |  |
Deep Agents Graph Memory | Graph context middleware and tools for Deep Agents that record and recall workflow findings using embedded LadybugDB. | TahaK29/deepagents-graph-memory |  |
Tenuo | Task-scoped authorization for LangChain agents that limits each task to the tools, argument values and time it needs, enforced before execution, with delegated agents receiving a subset of the parent’s permissions. | tenuo-ai/tenuo |  |
ISNAD | Claim-level provenance and trust grading for agent outputs. Grades every transmitter in a claim’s chain (source, scraper, model) in a living registry, caps the chain at its weakest link, and quarantines fabricated (mawḍūʿ) chains—with a tamper-evident audit trail. | alizahidraja/isnad |  |
attenu-guard | Per-agent permission enforcement on every tool call and subagent handoff through wrap_tool_call (a subagent’s permissions are a computed subset of its parent’s), with a hash-chained audit log verified offline | attenu-io/attenu-guard |  |
PIIAnonymizationMiddleware | Protect personal data (PII) in prompts, tool calls, and agent conversations. Hides sensitive values from the model, then restores the real values in the reply and at the tool boundary, so tools and the user still get the real data. Pluggable regex, NER, and LLM detectors, thread-consistent tokens, and token-by-token streaming restoration. | Athroniaeth/piighost |  |
Highflame | Runtime AI security guardrails (prompt injection, PII/DLP, content safety) applied as middleware via Highflame Shield (OWASP LLM Top 10). | highflame-ai/highflame-sdk |  |
LLMSlimToolOutputMiddleware | Agent middleware that shortens long tool results once, before the model first sees them, so later model calls resend identical bytes. Local, no model call. | Thanatos9404/llmslim |  |
TracekitMiddleware | Tracekit asks a separate signer to allow, deny or hold each agent tool call for a person before it runs, and records every call and result in a signed, hash-chained log that can be verified offline. | Cygnux-Labs/Tracekit |  |
Vulnify | Runtime authorization for tool calls with ALLOW, REVIEW, or BLOCK decisions, risk scores, optional human approval, and error ToolMessages for blocked calls. | vulnify/vulnify-sdk-python |  |
xaidr | In-process runtime security sensor for AI agents. LangChain middleware covering input, tool-call, and output boundaries for prompt injection, jailbreak, DLP, and A2A-protocol scanning. | delphisecurity/xaidr |  |
prompt-shield | Runtime prompt-injection firewall. Scans inputs, tool results, and outputs across detectors and output scanners with block, flag, and log modes. | mthamil107/prompt-shield |  |
Cisco AI Defense | Runtime security inspection | cisco-ai-defense/ai-defense-langchain-middleware |  |
compact-middleware | Claude Code’s compaction engine as LangChain middleware. Multi-level context compaction for long-running agents. | emanueleielo/compact-middleware |  |
SRARiskGateMiddleware | Checks stablecoin payment tool calls before they run and holds or rejects those that break spend, recipient, or asset rules. | sriram1983007-dev/langchain-sra-riskgate |  |
SeekritCredentials | Scopes secrets per model call and per tool call. Credentials stay as placeholders in agent code and are substituted at the HTTP boundary against a default-deny allowlist, so an unlisted tool cannot reach a credential. | seekritdev/python-sdk |  |
OpenBox | Real-time governance for LangGraph and Deep Agents. Policies, guardrails, HITL, OTel hook governance, and behavior rules. | OpenBox-AI/openbox-langgraph-sdk-python |  |
ScaleDown middleware | Conversation summarization, context and tool result compression, and structured extraction | scaledown-team/langchain-scaledown |  |
NeuralTrust TrustGuard | TrustGuard evaluate middleware. Allow, block, report, or transform agent input, model output, and optional tool traffic. | NeuralTrust/langchain-neuraltrust |  |
Nuggets Authority | Pre-execution authority enforcement for tool calls. Verifies a scoped, signed delegation before each tool runs, fails closed on deny, and emits an independently verifiable cryptographic proof of every decision. | NuggetsLtd/langchain-nuggets |  |
GoAT | langgraph-goat records candidate plans, scores, and why each was chosen or rejected as a LangGraph agent plans, then renders a natural-language explanation for debugging and end-user transparency. | Samsung/langgraph-goat |  |
Pydantic Monty Code Interpreter | Provides Code Interpretation inside Pydantic’s Monty Runtime | langchain-monty |  |
ProvaredMiddleware | Checks each tool call through wrap_tool_call against a passkey-signed permission, runs the tool only if allowed, and writes a signed, hash-chained receipt that can be verified offline. | provared/provared |  |
langchain-runcycles | Pre-execution budget authority for model calls, tool calls, and agent loops | runcycles/langchain-runcycles |  |
PromptFirewallHandler | Sub-millisecond PII detection and prompt injection firewall for LangChain. Zero network, zero GPU, pure Rust core via PyO3. | TimurRakhmatullin86/langchain-promptfirewall |  |
TealTiger | Deterministic governance middleware. Policy enforcement, cost limits, tool allowlisting, FREEZE rules, and SARIF audit evidence with no LLM in the governance path. | agentguard-ai/tealtiger |  |
langchain-router | Phase-based model routing. Routes execution turns to a fast model, keeps the primary for planning and recovery. | johanity/langchain-router |  |
langchain-task-steering | Implicit state-machine middleware for ordered task pipelines with per-task tool scoping, dynamic prompt injection, and composable completion validation. | edvinhallvaxhiu/langchain-task-steering |  |
BastionGuardrailMiddleware | Prompt-injection and jailbreak detection. Screens user input and tool results, including indirect injection through retrieved content, before the model runs. Runs locally on your own hardware. | bastion-soft/bastion-prompt-protection |  |
DynamicToolSelectorMiddleware | Drop-in agent middleware providing sub-100ms hybrid dense and sparse vector tool retrieval over zvec. Eliminates prompt token overhead by >80% and outperforms LLM-based tool selection on Berkeley Function Calling Leaderboard (BFCL) benchmarks. | RauhanAhmed/langchain-dynamic-tools-middleware |  |
ATLTaskMiddleware | Automatic task-time routing to executable MCP and provider capabilities, with failure redecision and outcome reporting. | wenhua6666668-oss/langchain-agenttrafficlab |  |
advisor-middleware | Claude Code’s advisor pattern as LangChain middleware. Pairs a fast executor model with a powerful advisor model that intervenes only on critical decisions. | emanueleielo/advisor-middleware |  |
TurnWindowMiddleware | Cross-provider LangChain agent middleware that retains the most recent human-anchored conversation turns before each model call while keeping tool-call and tool-result messages together. | bharatkhanna-dev/langchain-turn-window |  |
MeragptDecider | Typed decisions (yes/no, choice, score) with calibrated probabilities from meraGPT Decider 1, as a Runnable, a tool, and agent middleware. | Sandeep85Das/langchain-meragpt |  |
Wity | Fixed-option questions about text in one pass, with a probability per option, for model routing, tool-risk checks, and classification. | info-alphanimble/langchain-wity |  |
LoopGrid | Native LangChain AgentMiddleware for signed decision evidence, policy enforcement, human approval, and guarded tool execution through LoopGrid Core. | loopgridio/loopgrid-langchain |  |
RelayShield | Mandatory pre-execution gate that blocks connect_mcp_server and install_mcp_package tool calls when RelayShield reports risk. | nzdsf2-gif/langchain-relayshield |  |
RecallChatMessageHistory | Chat message history stored in a Recall session, so Recall also learns about the user from it. | recallmem/recall |  |
RecallMiddleware | Long-term memory for create_agent agents: adds what Recall learned about the user to the system prompt and records each exchange. | recallmem/recall | |
MaskflowReversibleAnonymizer | Reversible PII anonymizer / deanonymizer, a drop-in for langchain-experimental’s Presidio anonymizer (same .anonymize / .deanonymize / .deanonymizer_mapping and JSON/YAML save-load). Adds the Indian identifiers most tools miss (Aadhaar, PAN, GSTIN, UPI, IFSC, ABHA, names, addresses), a streaming-aware deanonymizer Runnable, and an optional leak-guard callback that fails a call closed if PII reaches the model. | maskflow/maskflow |  |
langchain-collapse | Preventive context management. Collapses consecutive tool-call groups before they fill the context window. | johanity/langchain-collapse |  |
Tessera | Signed trust labels and taint-tracking that gate tool calls when context contains untrusted segments. | kenithphilip/Tessera |  |
CTRLRunMiddleware | Policy-gated tool calls through wrap_tool_call: a refused call never reaches the tool and the model is told which rule refused it; an effect key executes at most once across processes sharing a store; a tool that raises leaves the effect unresolved rather than retried; an approve decision holds the call for a person; a receipt for every decision, refusals included. | CTRLRun/ctrlrun |  |
DeepKeep middleware | DeepKeep AI Firewall pre-moderation and post-moderation for LangChain agents. |  | |
AxioRank | Security gateway for AI agents: score tool calls and model turns against policy with allow, deny, and redact. | AxioRank/langchain-axiorank |  |
SubmilliMiddleware | Code-mode middleware that runs agent-written TypeScript programs on submilli-server under a policy and can call the agent’s own tools. | submilli/submilli-python |  |
langchain-pop | Portable persona loading, legacy migration, and boundary-aware tool filtering for LangChain agents. | joy7758/persona-object-protocol |  |
OWASP Agent Memory Guard | Runtime defense against AI agent memory poisoning (OWASP ASI06). Scans messages, model responses, and tool outputs locally with block, warn, and strip modes. | OWASP/www-project-agent-memory-guard |  |
langchain-distil | Reversible, certified context compression. Digests large tool outputs and message history before the model call (tool and function messages reversibly, human and system losslessly, the model messages never rewritten), with byte-exact recovery of every digest. Decision-equivalence between compressed and full context is certified offline by a statistical non-inferiority gate. | dshakes/distil |  |
comply54 | Runtime compliance enforcement for AI agents under African data protection and financial-sector regulations (deny, escalate, audit, or allow). | comply54/langchain-comply54 |  |
DNS-AID | DNS-based agent discovery via the DNS-AID protocol. Auto-publishes agents on startup, auto-unpublishes on shutdown, and provides discovery tools. | IngmarVG-IB/langchain-dns-aid |  |
NoPII | Runtime PII tokenization. Detects personal data in outbound prompts, replaces it with deterministic vault tokens before the request reaches the LLM, and restores the original values in the response. | Enigma-Vault/NoPII |  |
text2sql-framework | Replaces RAG with recursive tool use: the agent explores, writes, tests, and self-corrects using one execute_sql tool. | Text2SqlAgent/text2sql-framework |  |
langgraph-state-machine | Section-based flow control for LangGraph React agents. Divides conversations into discrete phases with scoped tools, prompts, auto-transitions, branching, and optional per-section LLM override. | mahmoud661/langgraph-state-machine |  |
VecrContextCompressor | Deterministic LLM context compression with a regex retention whitelist for structured tokens before the model call. | h2cker/vecr |  |
ACGSMiddleware | Checks each agent tool call against ACGS constitution rules before it runs; blocked calls are denied, and review rules pause the agent. | acgs-ai/langchain-acgs | N/A |
ATR Guardrail | Runtime detection of prompt injection, tool poisoning, and unsafe tool calls using Agent Threat Rules. Halts the agent or blocks the tool call on a critical finding and keeps an audit trail. | Agent-Threat-Rule/agent-threat-rules | N/A |
eager-tools | Reduces agent wall-clock latency by dispatching each tool call the moment its streaming block closes, overlapping tool execution with LLM generation. | cloudthinker-ai/eager-tools | N/A |
Haldir | Governance layer for LangChain agents with scoped sessions, encrypted secrets, hash-chained audit, and policy enforcement. | ExposureGuard/haldir | N/A |
NVIDIA middleware | Model routing and Nemotron 3 Ultra harness optimization | langchain-ai/langchain-nvidia, langchain-ai/deepagents | N/A |
如果您想贡献集成，请参考 贡献集成。
官方原文：src/oss/python/integrations/middleware/index.mdx · 非官方翻译，译文由 AI 生成，以英文原文为准。