Agents & developers

A complete local agent platform
with open interfaces.

OpenMiRA ships with an agent runtime, long-term memory and human review, and opens the same capabilities to external agents and apps over MCP and HTTP API. Local models by default, cloud models optional. Data stays on the device, and operations with side effects are confirmed by a person.

01 · Architecture

An agent platform with open interfaces.

OpenMiRA ships with complete agents and apps; the same interface is open to external agents, and models are swappable. Your data, memory and review records stay on your own device.

Callers
Agents · Apps
Claude Code · connectedMCP-capable agentsYour own apps (HTTP API)OpenMiRA app (same interface)
MCP · HTTP
Agent platform
Agent platform
OpenMiRAData and memory stay on the device
Agent runtime
Planning, tool calls and long-running tasks; the bundled OpenMiRA app is built on it
Memory
Long-term multimodal memory: people, places, events and preferences, linking the same person across years
Search & sources
One search across documents, images, audio and video; every answer cites its source
People review
Operations with side effects run only after a person confirms
Outbound ledger
Every cloud call is logged and exportable
Local by default · cloud optional
Models
Inference
Local models: llama.cpp · MLX · ComfyUICloud models (optional): OpenAI-compatible providers
Runs on
Your hardware
x86ARMRISC-VPreview
Scope
Every stage of the loop
is callable by your agents.
Here is what’s open today; anything with side effects follows “read-only by default, writes need human confirmation.”
Stage of the loopAPIWhat agents can callStatus
Library · Knowledge baseMCP · HTTP APISearch, cited Q&A, reading filesAvailable
CreationHTTP APIImage and video generation; lengths the machine can’t finish are refused up frontAvailable
CreationMCPCalled through “proposal → human approval”; approved actions run with exactly the proposed parametersPlanned
Publishing—External agents submit a publish request; a person approvesPlanned
02 · Integration

Three ways to integrate.

01 · Agents
Connect over MCP
Agents that support MCP (such as Claude Code) can call search, cited Q&A and the library. The tool list is synced from the server at startup; tools with side effects (delete, send out, modify data) are not exposed over MCP.
Claude Code
# The token is in the token file in the data directory
claude mcp add --env OWNWARD_TOKEN=$TOKEN ownward -- ownward-mcp \
  --url http://127.0.0.1:8090
02 · Apps
Connect over the HTTP API
Search, Q&A, long-running tasks and generation share one local interface; streaming results use SSE. Every request must carry a token.
Terminal
curl http://127.0.0.1:8090/api/v1/search \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"query": "payment terms in the supplier contract"}'
03 · Scenarios
Scenario packs
One JSON file: a few parameters to fill in, plus a preset prompt. Installing it adds a scenario, and you can read the full contents before installing.
Scenario catalog
# The scenario catalog and each pack are plain JSON files
GET /hub/index.json
GET /hub/packs/<id>.json
03 · Models

Local by default. Cloud if you want it.

Every task has a local default model. Whether to use the cloud, and which provider, is up to you. When cloud chat is on, automatically extracted memories are not sent to the cloud.

What it doesDefaultOptionalStatus
Chat & planningLocal chat modelsA cloud chat provider (OpenAI-compatible), switching the whole conversationAvailable
Recognizing public figuresOfflineA cloud vision model, called only when you trigger “Look it up online”Available
Images & videoGenerated locally—Available
Per-task routing—Pick local or cloud separately for chat, planning and vision; configure several providersPlanned
Compatibility
Local is the default; cloud is optional. The lists are labeled Available, To verify or Planned.
Local models and inference engines
Default · all of it runs on the machine
Built inAvailable
llama.cppMLXONNX RuntimeComfyUI
External local servicesPlanned
OllamaLM StudiovLLMSGLangLocalAIJanXinference
Built-in engines cover chat, vision, embeddings, reranking and image and video generation. External local services are the model servers you already run, connected through an OpenAI-compatible interface.
Cloud model providers
Optional · switched on by you, task by task
InternationalTo verify
OpenAIAnthropicGoogle GeminixAIMistralCoherePerplexity
ChinaTo verify
DeepSeekQwen (Alibaba Cloud)Kimi (Moonshot AI)GLM (Zhipu)Doubao (Volcano Engine)MiniMaxERNIE (Baidu)Hunyuan (Tencent)StepFun
Aggregators and hostsTo verify
OpenRouterSiliconFlowTogether AIGroqFireworksAzure OpenAI
Through an OpenAI-compatible interface: enter the endpoint and a key. How compatible each one is depends on its own documentation, and we haven’t verified them one by one; preset templates are planned.
Agents and clients
Over MCP · HTTP API
VerifiedVerified
Claude Code
MCP standardTo verify
CodexCursorWindsurfGemini CLIVS CodeClineZedClaude DesktopOpenClawHermes AgentGoose
Only Claude Code has been tested end to end; the rest connect through the MCP standard and haven’t been verified one by one. Your own apps use the HTTP API.
04 · Boundaries

Read-only by default,
writes need human confirmation.

The more an agent is allowed to do, the more a mistake costs. So access is read-only by default, operations with side effects need a person’s confirmation, and everything sent out is logged.

Read-only by default
Agents connected over MCP can only call tools without side effects: search, Q&A and reading files.
Available
Writes need human confirmation
Tools that delete, send out or modify data are always refused for external callers. They run only from the workbench, after a person has checked the real parameters and confirmed.
Available
Outbound logging
Every cloud call records the provider, model, bytes sent and whether it succeeded; export it as CSV.
Available
External agents submit requests; a person approves
An agent submits the operation it wants to run along with its full parameters; it waits in OpenMiRA for a person’s approval, then runs with exactly those parameters.
Planned
Block outbound data by category
Memories, faces and voiceprints, and original files stay on the device by default; everything else follows your rules, and anything blocked is flagged clearly.
Planned

Join the developer preview.

Get the preview docs and samples; you’ll be among the first to hear when the integration options and planned capabilities go live.

Read the docsScenario catalog