How AgentLocal works
The local platform where knowledge engineers train AI assistants with skills, MCP servers, plugins, and markdown context — on simulated LLMs, before they ship. One docker-compose, full local stack.
What it is
AgentLocal is the local platform where knowledge engineers train their agent with skills, MCP servers, plugins, and markdown context — against wire-compatible simulated LLMs (OpenAI, Anthropic, Bedrock) — composed with battle-tested AWS emulators under the hood.
Why AgentLocal
Authoring and testing skills, MCP servers, plugins, and markdown context against simulated LLMs — instead of real provider APIs — collapses three problems at once.
01 · The cost
Iterate on skills, MCP servers, plugins, and markdown context all day. Test against simulated LLMs that match OpenAI, Anthropic, and Bedrock wire format. Real output via local LLM, recorded cassette, or scripted YAML. Never burn a real API token on dev work.
02 · The stack
Wire-compatible LLM simulation, the four artifact types, and composed AWS emulation (S3, Lambda, DynamoDB, OpenSearch) — all in one docker-compose. No vendor cloud. No phone-home. Air-gap deployable for regulated industries.
03 · The visibility
See what context the agent loaded, what skill it picked, what tool it called, what document it grounded on. Step through any agent loop. Diff sessions. Replay from any step. Debugging stops being guesswork.
The problem
Tweak a SKILL.md, an MCP server config, a tool schema, or a context document — and you've just paid in tokens, lost visibility into what the agent did, hoped it still works in production, and (if you're in a regulated industry) maybe broken a data-sovereignty rule.
Iteration friction
$0.10–$2 / test
Every change to a skill, MCP server, or context document costs real LLM tokens and 3–30 seconds to test. The loop is slow and expensive by default.
Context blindness
Black box
You can't see what context the agent loaded, what skill it picked, what tool it called, what document it grounded on. Debugging is guesswork.
Deployment gap
Laptop ≠ prod
Skills work on the prompt engineer's laptop, then break in production — different context loader, different model, different MCP behavior.
Data sovereignty
Banned
Regulated industries cannot send proprietary skills, MCP configs, or context documents to third-party APIs even for testing.
How it works
Point your existing SDK at AgentLocal. Nothing else in your code changes — same client, same model names, same parameters.
Before
# Real provider. Real cost. Real latency. export OPENAI_BASE_URL=https://api.openai.com export AWS_ENDPOINT_URL=https://aws.amazon.com
After
# Fully local. Zero cost. Sub-second. export OPENAI_BASE_URL=http://localhost:4577 export AWS_ENDPOINT_URL=http://localhost:4566
The composition
AgentLocal — the platform
LLM simulation · Skills · MCP servers · Plugins · Markdown context · Marketplace
OpenAI, Anthropic, Bedrock wire-compatible. Three backends per request: local LLM, recorded replay, or scripted. Author and test the four artifact types locally.
AWS emulator
S3, Lambda, DynamoDB, SQS, IAM, OpenSearch
under the hood
Your AI calls hit AgentLocal. Your AWS calls hit a battle-tested emulator. Same docker-compose.
What's in the platform
OpenAI, Anthropic, Bedrock wire-compatible. Three backends per request: local model, recorded replay, scripted YAML.
Author SKILL.md files locally. Test against simulated LLMs. Inspect what the agent does with your skill. Iterate at zero cost.
Develop and test MCP servers locally. Author plugin schemas for all three providers. Drop markdown context and watch the agent ground on it.
See what context the agent loaded, what skill it picked, what tool it called, what document it grounded on. Step through any agent loop.
Publish skills, MCP servers, plugins, and markdown bundles to the public marketplace or your team's private marketplace. Discover, install, version.
Score skill, MCP, plugin, and context quality against labeled cassette suites. Regression detection across versions. CI gating.
How we compare
LocalAI is an excellent open-source AI inference engine — it lets you run models on your own hardware with an OpenAI-compatible API. If your goal is "serve models locally," LocalAI is a great choice, and you can even use it as the LLM backend inside AgentLocal.
AgentLocal is a different layer of the stack. We're not an inference engine — we're a local platform for AI skill development. LocalAI runs models. AgentLocal is where you teach those models to actually do your work — by authoring and testing skills, MCP servers, plugins, and markdown context locally.
LocalAI
An inference engine. Optimized for serving real AI workloads — 36+ model backends, hardware acceleration, multi-user auth, model gallery, fine-tuning, distributed inference.
Best for: replacing OpenAI's hosted API with a local equivalent for production or sustained inference.
A local skill development platform. Author and test skills, MCP servers, plugins, and markdown context. Three backends per request (local LLM, recorded cassette, scripted YAML), full agent inspection, marketplace distribution.
Best for: knowledge engineers and AI teams training domain-specific AI assistants with proprietary content, tools, and processes — where iteration speed, inspection, and data sovereignty matter.
Frequently asked
v0.1 ships in Q3 2026. Design partners get earlier access — typically 2–3 months before public release. Subsequent releases follow a roughly quarterly cadence through v0.5 (marketplace) and v1.0 (eval platform) in Q2 2028.
LocalAI is an inference engine — it runs models locally with an OpenAI-compatible API. AgentLocal is a development platform — three backends per request (local LLM, recorded cassette replay, scripted YAML), Terraform/CDK + AWS composition, and an agent debugger. The two solve different problems and work well together: LocalAI can be the local-LLM backend inside AgentLocal. See the full comparison →
No. AgentLocal is for development, CI, and demos. Production calls go to real providers. The integration is one environment variable, so switching between local and live is trivial.
We are currently in our pilot phase and actively onboarding early users, developers, and design partners. During this period, the Community tier is completely free, allowing individuals and teams to build, test, and develop on the platform at no cost while we refine the product together with our pilot community. We’ll be inviting selected users to participate in the pilot and gain early access to new capabilities and support. Planned pricing after general availability includes a Team tier at $899 per user per year with a 5-user minimum, while Enterprise pricing is custom starting at $25K per year and includes advanced features such as SSO, audit logging, air-gapped deployment options, and dedicated support.
Self-hosted only. AgentLocal runs on your laptop, your dev cluster, or your CI runners — never on AgentLocal-operated infrastructure. There's no vendor cloud, no phone-home, no prompt data leaving your network.
AgentLocal orchestrates the AI dev platform; the underlying AWS emulation is composed in from battle-tested open-source emulators. You get one docker-compose, full local stack, with everything wired up. The AgentLocal reference deployment includes the AWS emulator setup — you don't configure two products separately.
Yes — that's a primary design goal. AgentLocal is air-gap deployable, runs entirely on your infrastructure, and never phones home. SOC 2 Type II is targeted for year one of Enterprise sales. HIPAA-compatible architecturally; GDPR-friendly by design.
Ready when you are
Design partners get free Team tier for 12 months, direct line to the founder, and the earliest access to every release. Ten slots, closing the cohort by end of quarter.
