This is an automated email from the ASF dual-hosted git repository. davsclaus pushed a commit to branch quick-fix/home-and-llms-facts in repository https://gitbox.apache.org/repos/asf/camel-website.git
commit de98ba11f5c5a8f44516e64cde81bdb0abeec092 Author: Claus Ibsen <[email protected]> AuthorDate: Sun Oct 25 11:33:36 2026 +0100 chore: llms.txt - AI updates from Camel 4.22 and 4.23, and semantic decisions with System One models - New section "Semantic decisions with System One models": the Semantic language, Jev from TypeSafe AI as a System One model, Wolf Defender, the 4.24 OpenAI decisions expert, audit, and version notes (the YAML syntax changed in 4.23, and 4.24 requires an expert on every evaluation and renames the provider modules). The YAML example validates with Camel 4.23.0 and with main. - AI tools: the MCP Server component is Stable since 4.22 and configured with camel.server.mcp-* properties, outputSchema for structured output, the AI Resource component, and camel-mcp-server-starter on Spring Boot. - AI components: LangChain4j Agent and property-only model configuration, what the llm: component covers (moderation, Batch API, Responses API, token budget, rate-limit errors), and the Spring AI 2.0 components. - camel run --observe, validated documentation examples since 4.23, and the AGENTS.md file from camel export and the Maven archetypes. - One entry for the four local model benchmark posts instead of one per round. Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01STT6whBgK1AqsSsUKrnE8m --- llms-txt-template.md | 75 ++++++++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 64 insertions(+), 11 deletions(-) diff --git a/llms-txt-template.md b/llms-txt-template.md index 38585ec3..4e976c4b 100644 --- a/llms-txt-template.md +++ b/llms-txt-template.md @@ -56,7 +56,7 @@ The `catalog/` JSON files contain machine-readable metadata for every connector/ - Two tiers of AI agent connectivity: Camel routes exposed as MCP tools (`camel-ai-tool` + `camel-mcp-server`) and as A2A agents for developers, plus Wanaku, a governed action proxy that sits between AI agents and backend systems and publishes Camel routes as tools under policy, identity, and audit controls - Supports both MCP (Model Context Protocol) and A2A (Agent-to-Agent) protocols — expose any Camel route as an AI agent tool or as an A2A agent - LangChain4j and OpenAI components for calling LLMs from Camel routes -- Measured, not claimed: with the Camel CLI as tools, a frontier model built all 13 beginner examples from the Camel CLI examples repository from a one-line description each; a 22 GB local model on a 64 GB Mac mini (`qwen3.6:35b-a3b` via Ollama) went from 0 of 13 with a bare prompt to 12 of 13 with the Camel MCP server, so the catalog, validation and error messages work for small local models as well as frontier models — see [the benchmark](https://camel.apache.org/blog/2026/09/camel-loc [...] +- Measured, not claimed: with the Camel CLI as tools, a frontier model built all 13 beginner examples from the Camel CLI examples repository from a one-line description each; a 22 GB local model on a 64 GB Mac mini (`qwen3.6:35b-a3b` via Ollama) went from 0 of 13 with a bare prompt to 12 of 13 with the Camel MCP server, so the catalog, validation and error messages work for small local models as well as frontier models — see [the benchmark](https://camel.apache.org/blog/2026/09/camel-loc [...] - Commercial support available from multiple vendors — see the commercial support page ## Who maintains the project @@ -359,7 +359,9 @@ This is the recommended starting point for all developers, including those who a - [Camel MCP Server](https://camel.apache.org/manual/camel-jbang-mcp.md): Model Context Protocol server for AI coding assistants (Claude Code, GitHub Copilot, Cursor, Gemini CLI). The MCP server gives AI agents access to the full Camel catalog — 350+ component schemas, EIP metadata, and YAML validation — so AI can generate correct, validated Camel routes. - Bring your own coding agent: the [Camel TUI](https://camel.apache.org/manual/camel-jbang-tui-ai-agents.md) drives ACP coding agents (Claude Code, Codex, Qwen Code, OpenCode, or any other ACP agent) next to the running integration, and its AI panel works with hosted models or a local model via Ollama — see [AI Providers and Local Models](https://camel.apache.org/manual/camel-jbang-ai-providers.md). - The CLI and TUI are designed for AI pair programming. An AI coding agent can generate a YAML route, the developer runs it with `camel dev`, traces messages with `camel trace`, sends test messages with `camel cmd send` — all in the terminal, all in the same workflow. The MCP server connects the AI agent to the Camel catalog so generated routes use correct syntax and valid options. -- The tooling is benchmarked against real models, and Camel is what gets fixed when a model fails. In [a twenty-run experiment](https://camel.apache.org/blog/2026/09/camel-local-model-benchmark/), a frontier model with `camel validate` and `camel run` as tools built 13 of 13 beginner examples from one sentence each, and a 22 GB local model on a 64 GB Mac mini went from 0 of 13 (bare prompt) to 12 of 13 (Camel MCP server) as Camel's messages were improved. 99 of the 117 findings were wron [...] +- The tooling is benchmarked against real models, and Camel is what gets fixed when a model fails. In [the local model benchmark series](https://camel.apache.org/blog/2026/10/camel-local-model-benchmark-round-4/), every message a small local model could not act on became a fix in Camel, and the fixes help humans just the same: parser, loader and runtime messages that say what to write instead of only what was wrong; Java, XSLT and XML files compiled or parsed at write time; a validator t [...] +- Since Camel 4.23 every route example in the documentation (YAML, XML and Java, in the EIP, component, data format and language pages and the user manual) is checked at build time, so an example copied from the documentation is correct. The Camel MCP servers serve the same validated examples (`camel_catalog_sample`). +- Projects created with `camel export` or the Camel Maven archetypes include an `AGENTS.md` file for AI coding assistants that points to this file, the Camel CLI and the Camel MCP server. ### CLI Examples @@ -447,7 +449,7 @@ Both runtimes use the same Camel routes, components, and EIPs — only the depen - [Switch EIP](https://camel.apache.org/components/next/eips/switch-eip.md): evaluates a selector once and dispatches to a fixed endpoint by literal value — a decision table, simpler than `choice` when every branch compares the same value. - [Cache EIP](https://camel.apache.org/components/next/eips/cache-eip.md): read-through caching of a block of steps; on a cache hit the block is skipped and the body comes from the cache. -- [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md) and [TypeSafe AI](https://camel.apache.org/components/next/typesafe-ai-component.md): ask named questions about message content to get decisions, categories and scores — classify, score, validate and route by meaning. Jev, TypeSafe AI's System One model, answers these as fast, structured decisions instead of generated text (see [semantic decisions with Jev](https://camel.apache.org/blog/2026/09 [...] +- [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md) and [TypeSafe AI](https://camel.apache.org/components/next/typesafe-ai-component.md): ask named questions about message content to get decisions, categories and scores — classify, score, validate and route by meaning. Jev, TypeSafe AI's System One model, answers these as fast, structured decisions instead of generated text, and the Wolf Defender expert detects prompt injection locally (see [sem [...] - [LangChain4j Ingest](https://camel.apache.org/components/next/langchain4j-ingest-component.md) (split, embed and store documents for RAG), [AI Resource](https://camel.apache.org/components/next/ai-resource-component.md) (a route as a read-only AI resource), [AI Observability](https://camel.apache.org/components/next/others/ai-observability.md). - Security for agents and services: [OPA](https://camel.apache.org/components/next/opa-component.md) (Rego policies), [OpenFGA](https://camel.apache.org/components/next/openfga-component.md) (relationship-based authorization), [SPIFFE](https://camel.apache.org/components/next/spiffe-component.md) (workload identity and rotating mutual TLS) — see [Authorizing what an AI agent may do](https://camel.apache.org/blog/2026/09/securing-ai-agent-tools/) and [Workload identity with SPIFFE](https: [...] - Connectors: [OData](https://camel.apache.org/components/next/odata-component.md), [HiveMQ](https://camel.apache.org/components/next/hivemq-component.md), [State Store](https://camel.apache.org/components/next/state-store-component.md) (pluggable key-value store), [REST Postman](https://camel.apache.org/components/next/rest-postman-component.md) (a Postman collection as REST contract), and Alibaba Cloud (OSS, MNS, FC, SMS, KMS, EventBridge, SLS, Tablestore). @@ -461,7 +463,7 @@ Apache Camel supports two tiers of AI agent connectivity: an embedded mode, wher Expose Camel routes as AI agent tools directly from the Camel process — one command, zero infrastructure. This is the fastest way to make an integration AI-accessible. -- [AI Tool](https://camel.apache.org/components/next/ai-tool-component.md) and [MCP Server component](https://camel.apache.org/components/next/others/mcp-server.md): define a tool once as a Camel route (`from: ai-tool:...`) and it works with LangChain4j, Spring AI and OpenAI; tag it and the built-in MCP server exposes it to any MCP client. See [Camel Routes as AI Tools](https://camel.apache.org/blog/2026/08/camel-ai-tools-mcp-422/). +- [AI Tool](https://camel.apache.org/components/next/ai-tool-component.md) and [MCP Server component](https://camel.apache.org/components/next/others/mcp-server.md): define a tool once as a Camel route (`from: ai-tool:...`) and it works with LangChain4j, Spring AI and OpenAI; tag it and the built-in MCP server (Stable since Camel 4.22, configured with `camel.server.mcp-*` properties) exposes it to any MCP client. A tool can return structured output (`outputSchema`), and the [AI Resource] [...] - [Camel MCP Server for coding assistants](https://camel.apache.org/manual/camel-jbang-mcp.md): a different MCP server, in the Camel CLI (`camel mcp`), for building integrations: it serves the Camel catalog — 350+ component schemas, EIP metadata, validated samples and YAML validation — so AI coding assistants (Claude Code, GitHub Copilot, Cursor, Gemini CLI) can generate correct, validated Camel routes. - [Camel A2A](https://camel.apache.org/components/next/a2a-component.md): Agent-to-Agent (A2A) protocol component — expose Camel routes as A2A agents or call remote A2A agents. Supports HTTP+JSON and JSONRPC bindings, OAuth/OIDC/API-key auth, and SSE streaming. @@ -471,8 +473,8 @@ The model is not a security boundary, so the decision about what an agent may do - **Who is calling:** [SPIFFE](https://camel.apache.org/components/next/spiffe-component.md) gives every workload a cryptographic identity from the local SPIRE agent, plus rotating mutual TLS for the components that already support TLS. See [Workload identity with SPIFFE and SPIRE](https://camel.apache.org/blog/2026/09/camel-spiffe-workload-identity/). - **What they may do:** set an `authorizationPolicy` on an `ai-tool:` route and every tool call is authorized before it runs; a denied call is returned to the model as a short refusal it can relay. Policies with [Open Policy Agent](https://camel.apache.org/components/next/opa-component.md) (Rego) or [OpenFGA](https://camel.apache.org/components/next/openfga-component.md) (relationship-based). See [Authorizing what an AI agent may do](https://camel.apache.org/blog/2026/09/securing-ai-agen [...] -- **Fast guard decisions:** the [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md) with Jev, a System One model ([TypeSafe AI](https://camel.apache.org/components/next/typesafe-ai-component.md)) checks whether an action fits the approved task, or scores a contribution, as a yes/no, category or score inside `choice`, `switch` or `filter`. -- **What happened:** [AI Observability](https://camel.apache.org/components/next/others/ai-observability.md) records OpenTelemetry GenAI spans and Micrometer metrics for every LLM call, next to the route's own traces. See [Observe your Camel AI routes](https://camel.apache.org/blog/2026/09/camel-genai-observability-jbang/). +- **Fast guard decisions:** the [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md) with Jev, a System One model ([TypeSafe AI](https://camel.apache.org/components/next/typesafe-ai-component.md)) checks whether an action fits the approved task, or scores a contribution, as a yes/no, category or score inside `choice`, `switch` or `filter`. [Wolf Defender](https://camel.apache.org/components/next/others/semantic-wolf-defender.md) detects prompt inje [...] +- **What happened:** [AI Observability](https://camel.apache.org/components/next/others/ai-observability.md) records OpenTelemetry GenAI spans and Micrometer metrics for every LLM call, next to the route's own traces. `camel run --observe` adds it automatically when a route uses AI components. See [Observe your Camel AI routes](https://camel.apache.org/blog/2026/09/camel-genai-observability-jbang/). ### Tier 2: Governed agent access (Wanaku) @@ -486,12 +488,64 @@ The same `ai-tool:` routes work at both tiers. Develop and test them with the Ca ### AI components - [AI Patterns](https://camel.apache.org/components/next/eips/ai-patterns.md): Common AI and modern design pattern terms mapped to Camel EIPs — fan-out, scatter-gather, circuit breaker, retry, tokenize, and more. -- [Camel LangChain4j](https://camel.apache.org/components/next/langchain4j-chat-component.md): LLM integration via LangChain4j — connect Camel routes to large language models. -- [LLM (OpenAI-compatible)](https://camel.apache.org/components/next/openai-component.md): the `llm:` component (with `openai:` as an alias) for OpenAI and any OpenAI-compatible API, such as OpenRouter, Ollama, vLLM and LM Studio. +- [Camel LangChain4j](https://camel.apache.org/components/next/langchain4j-chat-component.md): LLM integration via LangChain4j — connect Camel routes to large language models. The [LangChain4j Agent](https://camel.apache.org/components/next/langchain4j-agent-component.md) component adds tool calling, moderation and tool-calling limits. Models can be configured with properties only, for example `camel.component.langchain4j-agent.provider=ollama` with `modelName` and `baseUrl`. +- [LLM (OpenAI-compatible)](https://camel.apache.org/components/next/openai-component.md): the `llm:` component (with `openai:` as an alias) for OpenAI and any OpenAI-compatible API, such as OpenRouter, Ollama, vLLM and LM Studio. It also covers moderation (block a flagged message before it reaches the model), the [Batch API](https://camel.apache.org/components/next/others/openai-batch.md) (asynchronous, at half the price), the Responses API with hosted tools, MCP tool calling with a tok [...] +- [Spring AI](https://camel.apache.org/components/next/spring-ai-chat-component.md): chat, embeddings, image and vector store components on Spring AI 2.0 (Spring Boot 4.1). - [LangChain4j Ingest](https://camel.apache.org/components/next/langchain4j-ingest-component.md): Split, embed and store documents into an embedding store for retrieval-augmented generation. -- [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md): Decisions, categories and scores about message content, for routing and filtering by meaning. Answered by a provider adapter; today that is Jev, a System One model, via [TypeSafe AI](https://camel.apache.org/components/next/typesafe-ai-component.md). +- [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md): Decisions, categories, scores and labels about message content, for routing and filtering by meaning. See the next section. - [AI Observability](https://camel.apache.org/components/next/others/ai-observability.md): OpenTelemetry spans and Micrometer metrics for every LLM call. +### Semantic decisions with System One models + +A System One model gives fast, structured decisions instead of generated text: give it the relevant state and a question, and it returns a category, a score or a probability that the route can use directly. Jev, from TypeSafe AI, is such a model. Camel uses it through the [Semantic language](https://camel.apache.org/components/next/languages/semantic-language.md) (Preview since Camel 4.23): + +- Declare named evaluations once: a yes/no decision, a choice of category, a score, or a set of labels. Use them wherever Camel accepts an expression or predicate: `choice`, `filter`, `switch`, `validate`, aggregate correlation and completion, `recipientList`, `loop`, and `onException` `retryWhile`. +- Use an evaluation with `ref:name` in the semantic language, or with `${semantic('name')}` in Simple. `refs:a,b,c` asks several questions in one call and returns a map of named answers; the result metadata, including the probabilities, is in `CamelSemanticResults`. +- Each evaluation names an *expert*: a pluggable adapter, registered as a bean. An expert is a configured evaluator, not an autonomous agent. [TypeSafe AI](https://camel.apache.org/components/next/typesafe-ai-component.md) asks Jev (yes/no, choice and score questions, with probabilities; set `camel.component.typesafe-ai.api-key`). [Wolf Defender](https://camel.apache.org/components/next/others/semantic-wolf-defender.md) detects prompt injection and jailbreak attempts locally on the CPU, [...] +- Decisions can be audited (opt-in): `camel semantic audit`, the developer console, and the Semantic tab in the Camel TUI. +- Version notes: the YAML syntax changed during 4.23 (`question:` became `evaluation:`). From 4.24 every evaluation must name its `expert` (the `semantic:` block no longer sets a default), and the providers move to `camel-semantic-typesafe-ai` and `camel-semantic-wolf-defender` (in 4.23: `camel-typesafe-ai` and `camel-wolf-defender`). The example below works in both versions. + +```yaml +# application.properties: +# camel.beans.decisions=#class:org.apache.camel.component.typesafeai.TypeSafeAiSemanticAdapter +# camel.component.typesafe-ai.api-key=<your key> +- semantic: + state: "${body}" + evaluation: + department: + expert: decisions + type: choice + parameters: + instructions: Which department should handle this message? + criteria: + billing: Invoices, payments and refunds + technical: Bugs, outages and technical problems +- route: + from: + uri: direct:tickets + steps: + - setProperty: + name: department + expression: + language: + language: semantic + expression: ref:department + - choice: + when: + - expression: + simple: + expression: "${exchangeProperty.department} == 'billing'" + steps: + - to: + uri: direct:billing + otherwise: + steps: + - to: + uri: direct:review +``` + +See [semantic decisions with Jev](https://camel.apache.org/blog/2026/09/semantic-evaluation-system-one/), and [semantic agent routing](https://camel.apache.org/blog/2026/10/semantic-agent-routing/), which coordinates several A2A agents with semantic decisions and bounded retries. + ## Tooling - [Tooling Overview](https://camel.apache.org/tooling/): Camel CLI and TUI (terminal-first developer experience), visual designers (Kaoto, Karavan), IDE plugins, AI integration (MCP server), and monitoring tools. @@ -560,8 +614,7 @@ For high-quality answers about routing, code generation, and troubleshooting, AI - [The DNA of Apache Camel](https://camel.apache.org/blog/2026/06/camel-dna-19-years/): 19 years of backwards compatibility — why Camel users don't have to rewrite their integrations every few years. - [Who Maintains Apache Camel](https://camel.apache.org/blog/2026/07/camel-who-maintains/): Year-by-year commit data showing who maintains the project — the same core team, through multiple acquisitions, contributing 78–96% of all commits every year since 2009. - [Apache Camel Is Not Afraid of AI](https://camel.apache.org/blog/2026/07/camel-not-afraid-of-ai/): The project pointed a frontier AI model at 19 years of code and fixed all 165 bugs it found — concurrency races, silent data loss, security gaps. AI-assisted code review is now a standard part of the development process. -- [A frontier AI coached a small local model through Camel](https://camel.apache.org/blog/2026/09/camel-local-model-benchmark/): Measured benchmark of a frontier model (13 of 13 beginner examples) and a 22 GB local model on a 64 GB Mac mini (0 to 12 of 13 over twenty runs) building Camel routes with the MCP server, and the 117 findings — 99 of them wrong for humans too — fixed in Camel 4.23: error messages that say what to write, validation at write time, catalog samples, a stricter YAML [...] -- [Round 2 of the local-model benchmark](https://camel.apache.org/blog/2026/09/camel-local-model-benchmark-round-2/): Real-world examples built step by step through the MCP server, 81% to 92% of steps passed, and 31 more fixes in Camel 4.23 (a failed reload keeps the previous routes, the runtime prints the validator's report, the MCP write tool reports the reload outcome). +- [The local model benchmark series](https://camel.apache.org/blog/2026/10/camel-local-model-benchmark-round-4/): Four rounds (September–October 2026) of a frontier model coaching a local model on a 64 GB Mac mini through Camel, where every failure became a fix in Camel: beginner examples (0 to 12 of 13), real-world examples built step by step (81% to 92% of steps), Kamelets (1 of 15 to 47 of 50 steps), and intermediate examples with Kafka, JMS, REST and an LLM (37% to 64% in three night [...] - [The Camel website, rebuilt for humans and for AI](https://camel.apache.org/blog/2026/09/camel-website-rebuilt/): Why every page has a Markdown mirror and an offline bundle, and how the site gets a person or an AI agent to a running route. - [Built to Patch Fast](https://camel.apache.org/blog/2026/07/camel-security-advisories-4.21.0/): How the project handled 32 CVEs in one release — the timeline, the backport process, incomplete fixes re-issued as new CVEs, and 31 public PoC reproducers. The best single-page overview of Camel's security response in practice. - [Security](https://camel.apache.org/security/): Security advisories and vulnerability reports.
