Public beta · Open source

A PRIVATE
AI AGENT
FOR LOCAL
WORK.

Yemaka makes small Ollama models genuinely useful. Local memory, document retrieval, deterministic routing, and approval-gated tools — all running on your machine. No account. No telemetry. No cloud by default.

Install beta View on GitHub See safety defaults
Web app CLI TUI SQLite memory + RAG No telemetry MIT licensed
YEMAKA — LOCAL SESSION
loopback:7727
Ollama first
LOW-RESOURCE LOCAL MODELS
Yemaka makes small installed models useful by routing work before the model acts. No silent model pulls, bundled weights, or cloud fallback by default.
RuntimeOllama local
Rolesdefault / low_memory / coding
Policymanual downloads
Model policy
The user chooses model names already present on the machine. Yemaka stores role mapping, not model artifacts.
Local context
SQLITE MEMORY + FTS5 RAG
Conversation memory, document chunks, tool runs, and workflow traces stay in local SQLite. FTS5 retrieval feeds bounded prompts.
Storelocal SQLite
SearchFTS5 RAG
Scopeworkspace grants
Owned context
No hosted vector database is required. Indexed notes and documents remain local and inspectable.
Policy gates
TYPED TOOLS + SAFE ACTION
Yemaka routes work through local tools, skills, and extensions. Reads stay bounded; writes and scheduled jobs require approval and leave audit trails.
Routingdeterministic lanes
Writessnapshot + diff
Trustmanifest + tests
Approval required
file:write
src/config.ts · 12 lines changed
Snapshot prepared · Diff ready to review · Rollback path confirmed
Shared core
WEB, CLI, TUI, LOCAL SERVER
The browser UI, command line, terminal UI, and loopback server all speak to the same Go core and local profile.
Web127.0.0.1:7727
CLIdoctor / ingest / chat
TUIterminal UI
Inspectable runtime
Doctor checks, heartbeat, replay traces, and QA review keep local operation visible from every surface.
Local first Runs from your machine. Loopback server, no cloud required.
Low resource Designed for small Ollama models and modest hardware.
Fully inspectable Typed tools, execution traces, diffs, logs, and QA review.
MIT licensed Open-source SiloTabs public beta. Own it completely.
System design

Small model.
Strong agent system.

Yemaka does the deterministic work before the model ever acts: retrieve context, scope the toolset, validate the plan, then verify the result.

Input
Installed Ollama model
Any model already installed. Yemaka does not pull models silently. You choose and control what runs.
+
System
Local memory, RAG, skills, and tools
SQLite-backed context, FTS5 document retrieval, skill packs, and approval-gated tool execution — entirely on your machine.
=
Result
Useful, auditable AI agent
A productive local agent with no external dependencies, no hidden data flows, and no surprise costs.
What it includes

A trusted core.
Practical surfaces.

chat + routing
Local chat with deterministic routing
Chat with installed local models while Yemaka routes each task into bounded, scoped execution lanes — no free-ranging agents.
memory + rag
SQLite memory and FTS5 document retrieval
Search local memory and indexed documents without a hosted vector database. All context stays in a single SQLite file you own.
approval gates
File writes require snapshot, diff, and approval
Every destructive action creates a full snapshot, displays a diff, waits for your approval, and logs the result with a rollback path.
skills
Reusable skill and domain packs
Local instruction sets help the agent perform repeated workflows reliably — loaded on demand, never silently active.
extensions
Agent-generated, sandbox-tested extensions
New capabilities are proposed, tested in isolation, registered explicitly, and kept outside the trusted core. Core policy is immutable.
diagnostics
Doctor, heartbeat, trace replay
Health checks, execution traces, tool logs, and QA review surfaces make failures visible, reproducible, and recoverable.
Protected defaults

Private until
you decide otherwise.

Yemaka starts from a clean local profile. Optional systems are explicit and policy-gated — nothing runs quietly in the background.

Default off
Cloud fallback disabled
Remote models are never used unless you explicitly configure them. There is no silent fallback to any hosted API.
Default off
Internet access disabled
Public web search requires an explicit provider setup and a policy gate. No background network calls.
Default off
Connectors disabled
External service adapters and APIs stay off until you configure and enable them individually.
Never
Zero telemetry
No remote account, no hosted dependency, no analytics. Nothing leaves your machine unless you route it there.
File writes
Snapshot before every edit
File changes are previewed as diffs and applied only after your explicit approval. Every write is logged and reversible.
Models
Manual downloads only
Yemaka lists your installed Ollama models. It does not download, update, or switch models without instruction.
Jobs
Approval before scheduling
Scheduled background jobs are created and activated only after you review and approve them. No silent crons.
Extensions
Tests required before trust
Generated code passes an isolated test suite before registration. Extensions cannot touch policy, secrets, audit, or rollback rules.
Interfaces

One local core.
Three surfaces.

Web app, CLI, and TUI all speak to the same Yemaka process. Switch between them without losing context.

web
Local web app
Runs on loopback at http://127.0.0.1:7727. Full session interface, accessible from any local browser.
cli
Command line
Inspect models, run doctor checks, chat, ingest documents, manage jobs, and configure policy gates.
tui
Terminal UI
A lightweight keyboard-first interface for constrained machines and terminal-native workflows.
yemaka — local session 
# list installed models and pick one
$ yemaka model list
  llama3.2:3b   ✓ installed
  qwen2.5:7b    ✓ installed
  phi3:mini     ✓ installed

# ingest local documents into SQLite FTS5
$ yemaka ingest ./docs
  indexed 47 files → yemaka.db

# run health checks before starting
$ yemaka doctor
  ✓ model reachable
  ✓ db healthy (WAL mode)
  ✓ tools registered
  ✓ approval policy loaded

# start a session — everything stays local
$ yemaka chat "Review this project"
  routing → planning → context fetch
  ⚠ file:write detected — approval required
Install

Public beta.
Built for local.

Install scripts prepare a Yemaka home directory, copy the web bundle if needed, and create a command shim. They do not download models, enable cloud, start background jobs, or activate internet access.

After install, yemaka doctor runs a full health check before you start your first session.

macOS / Linux
scripts/install/install.sh

yemaka doctor
yemaka serve
yemaka tui
Windows PowerShell
powershell -ExecutionPolicy Bypass `
  -File scripts/install/install.ps1

yemaka --help
yemaka doctor
example workflow 
$ yemaka model list
# selects an already-installed Ollama model

$ yemaka ingest ./docs
# local documents become searchable via SQLite FTS5

$ yemaka chat "Summarize this project"
# memory, RAG, tools, policy, and verification
# all remain entirely on your machine