Key Takeaways
- Plano (github.com/katanemo/plano) is a free, open-source, self-hosted AI proxy and data plane β not a hosted SaaS product
- Developed by Katanemo, and built on Envoy by Envoy's own core contributors
- Apache 2.0 licensed, confirmed via the GitHub repository's license file
- Four core capabilities per its own README: agent orchestration, model routing ("Model Agility"), zero-code observability ("Agentic Signals" plus OpenTelemetry), and guardrail/moderation Filter Chains
- Routing runs through a purpose-built, roughly 4-billion-parameter orchestrator model ("Plano-Orchestrator") rather than a general-purpose LLM or a rules-only framework
- Over 7,050 GitHub stars and 480+ forks as of this review
π In One Sentence
Plano is a free, open-source (Apache 2.0), Envoy-based AI-native proxy server and data plane, developed by Katanemo, with over 7,050 GitHub stars, that adds LLM routing, agent orchestration, observability, and guardrails to agentic applications outside your application code.
π¬ In Plain Terms
Plano is a self-hosted proxy you run in front of your AI agents. Instead of writing your own routing logic, telemetry, and safety filters into every agent, you declare your agents and model providers in a YAML config file, and Plano handles picking the right model, tracing what happened, and applying guardrails β for free, since Plano itself has no license fee.
πNote: This review is the deep-dive companion to Plano's entry in the Local LLM Software Directory β see that page for how Plano compares at a glance to dozens of other local and self-hosted AI tools.
What Is Plano?
Plano is a self-hosted, out-of-process proxy and data plane that sits between your agentic applications and the LLM providers or sub-agents they call, handling routing, orchestration, observability, and guardrails in one layer. Its own GitHub description calls it "an AI-native proxy server and data plane for agentic apps," aimed at pulling infrastructure concerns like routing and moderation out of application code and into a shared, configurable layer.
- Product type: a self-hosted proxy/data-plane you deploy yourself (Docker container or built binary), not a hosted API or SaaS gateway
- Developer: Katanemo, per the repository's GitHub organization
- Repository: github.com/katanemo/plano, created July 9, 2024
- License: Apache 2.0, per the GitHub repository's license file
- Foundation: built on Envoy, the widely used open-source edge/service proxy, by contributors who worked on Envoy itself, per Plano's own README
- Scale: over 7,050 GitHub stars, 480+ forks, and 140 open issues as of this review
Plano's Project History and Version Milestones
Plano's GitHub repository was created in July 2024, and the project has shipped continuously since, adding LLM routing, agent orchestration, and observability features under active development. Plano positions itself as backed by Katanemo's own LLM research, including the purpose-built orchestrator model used for its routing decisions.
- 1Repository created β July 9, 2024
Why it matters: Plano's canonical GitHub repository (katanemo/plano) was created, per GitHub's repository metadata. - 2v0.4.36 β August 19, 2026: Most recent tagged release as of this review
Why it matters: The latest published release on GitHub at the time of this review; check the [releases page](https://github.com/katanemo/plano/releases) directly for anything shipped after this review's publish date. - 3Continuous main-branch development
Why it matters: GitHub's repository metadata shows commits pushed to the main branch after the latest tagged release, indicating ongoing development beyond the most recent version tag β verify current release status before assuming feature parity with the latest commits.
What Does Plano Actually Do?
Plano's feature set centers on moving infrastructure concerns for agentic apps out of application code and into a shared, self-hosted proxy layer. Here is what each part actually does, per Plano's own GitHub README and documentation.
- Agent orchestration β low-latency routing between agents declared in a YAML config, without modifying application code; new agents can be added by editing config rather than rewriting routing logic
- Model routing ("Model Agility") β routes requests by explicit model name, by semantic alias, or automatically via configured preferences, so application code can call a logical model name rather than hardcoding a specific provider
- Agentic Signals β Plano's own term for zero-code capture of usage/behavior signals, plus OpenTelemetry-compatible traces and metrics emitted across every agent it proxies
- Guardrail and moderation Filter Chains β a configurable mechanism for building jailbreak protection, moderation policies, and memory-consistency behavior into the proxy layer rather than each individual agent
- Purpose-built routing model β routing decisions run through Plano's own roughly 4-billion-parameter "Plano-Orchestrator" model rather than a general-purpose LLM call or a purely rules-based router, per Plano's README
- Framework- and language-agnostic β agents are just HTTP servers implementing an OpenAI-compatible chat completions endpoint, so Plano works with agents written in any language or AI framework
- LLM gateway mode β Plano can also be used purely as an LLM routing gateway from application services, independent of its agent-orchestration features, per its own documentation
Usage Examples: Two Ways to Use Plano
These are concrete workflows built from Plano's own documented README example β not hypothetical use cases.
Install Plano
Plano installs free via Docker or a built CLI binary, and its source code is on GitHub. Per its own Quickstart Guide, install the prerequisites, then start Plano against a config file with planoai up config.yaml.
Source | Link |
|---|---|
| GitHub repository (source code, Apache 2.0) | github.com/katanemo/plano |
| Quickstart guide (install + prerequisites) | docs.planoai.dev/get_started/quickstart |
| Full documentation | docs.planoai.dev |
| Releases (tagged versions and prebuilt binaries) | github.com/katanemo/plano/releases |
Plano requires following the prerequisites guide before running planoai up config.yaml; there is no single-file GUI installer, since Plano is a proxy/data-plane component meant to run alongside your services rather than a standalone end-user application. Verify the current recommended install method on the quickstart guide before running any command, since install instructions can change between releases.
Plano Pricing: Is Plano Really Free?
Plano itself is free and open source (Apache 2.0), with no license fee for self-hosting it. The nuance is in what powers its routing: Plano and the Plano family of LLMs (such as Plano-Orchestrator) are hosted free of charge in the US-central region by Katanemo, explicitly to give developers a first-run experience without setup β per the project's own README, that free hosted tier is intended for development, not production scale.
- Plano's own software: free, open source, Apache 2.0 licensed, no paid tier for the proxy itself
- Plano-family routing/orchestrator models: hosted free in US-central for a first-run developer experience, per Plano's own README
- To run at production scale, Katanemo's README states you must either run these routing models locally yourself, or contact Katanemo via Discord for hosted API keys β no public self-serve pricing page for that hosted-key option was found for this review
- LLM provider costs are separate: routing requests to OpenAI, Anthropic, or other providers through Plano still costs whatever that provider charges β Plano does not itself add a markup, per its role as a routing layer rather than a model provider
Plano vs. LiteLLM
Plano and LiteLLM are both self-hosted layers for managing LLM traffic, but they start from different premises: LiteLLM is primarily a unified API/SDK and proxy for calling 100+ LLM providers through one OpenAI-compatible interface, while Plano is built as an Envoy-based data plane with agent orchestration as a first-class feature, not just model routing.
Core foundation
- Plano vs. LiteLLM:
- Plano is built on the Envoy proxy; LiteLLM is a Python-native SDK and proxy server, not Envoy-based.
Agent orchestration
- Plano vs. LiteLLM:
- Plano treats multi-agent routing as a core, first-class feature via YAML-declared agents; LiteLLM focuses primarily on LLM/provider routing rather than agent-to-agent orchestration.
Routing mechanism
- Plano vs. LiteLLM:
- Plano uses a dedicated ~4B-parameter orchestrator model for intent-based routing; LiteLLM routes primarily via explicit config rules, load balancing, and fallback lists.
Observability
- Plano vs. LiteLLM:
- Plano calls its telemetry "Agentic Signals" plus OpenTelemetry; LiteLLM offers usage tracking, spend logs, and integrations with observability tools like Langfuse.
License
- Plano vs. LiteLLM:
- Both are Apache 2.0 for their core open-source offering, per each project's own repository.
Both tools are self-hosted and free at their core; the practical choice depends on whether you need Envoy-grade proxy infrastructure and native agent orchestration (Plano) or a simpler, provider-routing-focused gateway with a large provider-compatibility list (LiteLLM). Verify current feature parity on each project's own documentation before choosing, since both ship updates frequently.
Who Should Use Plano?
Whether Plano fits depends on whether you are building multi-agent applications that need shared infrastructure for routing, observability, and guardrails, rather than a single simple LLM call.
Competitors and Alternatives
Plano is most often compared to LiteLLM, AIClient2API, and Langfuse β its main differentiator is being an Envoy-based data plane with native multi-agent orchestration, rather than a purely application-layer routing library or an observability-only tool.
Tool | Best known for | Link |
|---|---|---|
| LiteLLM | Unified proxy/SDK for calling 100+ LLM providers through one OpenAI-compatible API | LiteLLM review |
| AIClient2API | Lightweight local proxy that converts various AI client protocols into unified API endpoints | AIClient2API review |
| Langfuse | Open-source LLM observability, tracing, and evaluation platform | Langfuse review |
This list reflects tools commonly compared to Plano in the routing/gateway/observability space, not an independent PromptQuorum ranking β see the Local LLM Software Directory for the full, regularly updated catalog, including Plano's own directory entry. Verify each tool's current pricing and feature set before choosing.
Common Mistakes When Evaluating Plano
Most confusion about Plano comes from its generic name, assumptions about it being a hosted SaaS, or unclear expectations about what "free" covers.
Frequently Asked Questions
What is Plano?
Plano (github.com/katanemo/plano) is a free, open-source (Apache 2.0), Envoy-based AI-native proxy server and data plane for agentic applications, developed by Katanemo, that handles LLM routing, agent orchestration, observability, and guardrails.
Is Plano free?
Yes, Plano itself is free and open source, with no license fee for self-hosting it. Katanemo also hosts Plano's own routing/orchestrator models free in the US-central region for a first-run developer experience; production-scale use of those models requires self-hosting them or getting hosted API keys via Discord.
How do I install Plano?
Follow Plano's own Quickstart Guide to install the prerequisites, then run planoai up config.yaml against a YAML config declaring your agents and model providers.
Is Plano a hosted service or self-hosted software?
Plano is self-hosted: you deploy it yourself as a Docker container or built binary alongside your own agents and services. Katanemo separately hosts Plano's own small orchestrator models for free at the development-usage level.
What does Plano actually route?
Plano routes LLM and agent traffic: it can route between multiple agents based on intent, and route LLM API calls to a specific model by name, semantic alias, or automatic preference-based selection.
Is Plano built on Envoy?
Yes, per its own README, Plano is built on the Envoy proxy by contributors who worked on Envoy itself, functioning as an out-of-process data plane rather than an in-process library.
Who develops Plano?
Katanemo develops Plano, per the GitHub organization hosting the katanemo/plano repository.
Does Plano work with any programming language or framework?
Yes β agents only need to implement an OpenAI-compatible chat completions HTTP endpoint, so Plano works with agents written in any language or AI framework, per its own documentation.
Does Plano require its own model to work?
Plano's routing decisions run through its own purpose-built roughly 4-billion-parameter orchestrator model, but the actual LLM responses your agents use can come from any configured provider (OpenAI, Anthropic, local models, and others) that Plano routes traffic to.
How is Plano different from LiteLLM?
Plano is an Envoy-based data plane with native multi-agent orchestration as a core feature; LiteLLM is a Python-native SDK/proxy focused primarily on unifying calls to 100+ LLM providers. See the Plano vs. LiteLLM comparison above for details.
Has PromptQuorum independently tested Plano's claims?
This review is based on Plano's own GitHub repository, README, and documentation, rather than hands-on benchmarking by PromptQuorum.