Conductor AI agents

ORCHESTRATING AGENTS · BACK ON TRACK GURU CLAUDE GPT GEMINI GROK GEMMA vLLM ··· AISHA
Back on Track
Alpha · v0.9 Self-managing AI orchestration platform
55.
MCP Tools
90+
n8n workflows
5.
tier model · adaptive
90%
target range
One-Click Start

Choose a class.
The train is already waiting on the platform.

Download the extension, mobile app, or desktop Workbench. Connect to our cloud or set up AISHA on your own—the code is publicly available on GitHub.

Safety · What Comes Out of Your Machine

Source code
It stays with you.

Extension sends Aishe only signals — diagnostics, file roles, Git branches, filenames, and the active language. No code content, no secrets in .env. The backend runs on on our private EU cloud. Verifiable in the repo:

tech_stack active_files git_branch recent_commits diagnostics dirty_files active_language

→ extensions/aisha-dirigent/src/aisha-context-sync.ts

With an account · learning your project

Aisha with You
The assignment will be approved.

After you log in, we'll be in touch with you passes the test, explore the possibilities, and fill in the gaps. Together, you gradually build Your area of expertise — and it will never overwrite anything it isn't supposed to. Without your confirmation, AISHA won't do anything—it will only make suggestions.

01 You go over the assignment together
02 Will explore options and recommend a course of action
03 Your documents to KB — always uploaded by you
04 Changes are always made with your approval
AI Security · OWASP Top 10

Testing
herself.

The entire pipeline runs in OWASP Top 10 mode—from injection to insufficient logging. It also heavily hardens the AI layer: consent gate, rate limiting, input validation, JWT scope, audit trail. No binary content is fed into the model—only the SHA-256 hash and edited text fragments.

RLS SECURITY DEFINER RPC audit_journal consent gate rate limiting CORS allowlist editorial staff handling sensitive data size cap

→ docs/security/SECURITY.md · 10/10 rating

EU Private Cloud conductor.aisha.guru
SSO Keycloak · OIDC
Status ● all systems green
13,000CZK
Conductor daily rate · 4-hour block
40+applications
The entire stack · one command
9,000+
Inspections in the CI Gate Suite
6languages
Basic Localization
−50%
Batch Queue · OpenAI & Anthropic
What AISHA Does

He coordinates the agents.
He puts them back on the track.

01
AISHA, conductor

The master orchestrator reads signals from 14 channels, decides on the playbook, and delegates tasks to sub-agents.

02
Agents Aligned

Copilot, Cursor, and Claude will share a common knowledge base. 55 MCP tools. One context.

03
Back on Track

Compliance Gate, Copilot Watcher, and Self-Healing bring drift back into alignment—proactively, not reactively.

04
One goal, maximum impact

Story-level delivery. Planning, code review, testing, deployment—one track, one direction.

Use it however you want

Four modes.
One brain.

AISHA adapts to your way of working. Call models directly through the LLM Gateway, let SoulForge shorten your prompts, isolate tasks in Exec sandboxes, or launch an OpenClaw network of agents—which AISHA orchestrates for you.

01 own keys · own accounts

LLM
Gateway

Proprietary API keys for OpenAI, Anthropic, Google, xAI, and Mistral. Proprietary batch endpoint. Proprietary self-hosted vLLM with Qwen3 on your GPU. AISHA simply routes, retries, and measures. No middleman on the receipt—you pay the provider directly.

Bring Your Own Key batch API self-hosted vLLM Maestro
// custom key, custom account
gateway.use({ provider: "openai",
  key: process.env.OPENAI_KEY })
02 token savings

Seoul
Forge

It sends shorter prompts. It summarizes the history, selects only relevant KB chunks, and deduplicates the context. When a query is reduced from 8,000 tokens to 1,200, you save 6×.

prompt squeeze context select dedup summarize
// fewer tokens, same response
forge.compress(ctx, { target: 1,200,
  keep: "hits" })
03 isolated environment

Exec
sandbox

The plugin code runs in a Docker container with an explicit fetch allowlist, a KV store, and an LLM proxy. There are no side effects from the agent on the host system.

node:vm Docker RPC whitelist
// safe exec
exec.run(plugin, ctx, {
  timeout: 30,000 })
04 network of agents · managed by AISHA

Open
Claw

A network of specialized agents who pass tasks to one another, share context, and complement each other. AISHA oversees them, orchestrates their work, evaluates their performance, and escalates issues. No “autopilot”—just collective intelligence under a single conductor.

mesh agent handoff delegation audit
// "Divide it up among yourselves"
claw.dispatch(task,
  { orchestrator: "aisha" })
AISHA's Brain

Aisha IS
conductor.

It’s not a chatbot. When it’s thinking, it draws on four layers of knowledge—rules, memories, a long-term database, and conversational guidance. If any of these layers “doesn’t know” something, it won’t remain silent. It will ask you a question and seek your advice.

It is based on a preset

Aisha offers a universal foundation—ethics, safety, and planning. We'll go over the assignment with you, explore the possibilities, and fill in the gaps. Together, you’ll decide what works and what doesn’t. The best part is that your expertise grows with every project.

TAO
Identity & Will

Aisha's inner compass—who she is, what she's allowed to do, what she wants. Ethical boundaries, compliance gates, principles of collaboration. She works with a knowledge base, but it is not identical to it.

Hippocampus
Short-Term & Episodic Memory

ai_session_memory, conversation ring buffer, cross-session messaging. It remembers what we talked about and what came of it.

Ragnarok
Knowledge Base · Long-Term Memory

Alquist Insight RAG · BM25 + KNN hybrid · pgvector. Your documents, decisions, expert rules — everything Aisha “knows.” The query returns only what’s worth reading.

Maestro
Advanced Chat · Discussion with You

Alquist multi-turn — keeps the conversation flowing, with one turn leading naturally into the next. He knows when to speak, when to listen, and when to ask a question. They'll handle it with you here, not in my head.

Guidance · 4 Levels

The voice that
It will help you.

AISHA adapts to how much you know. Junior will get teacher tone, expert on silence supervisor. You switch with a single selection—the entire platform's behavior shifts with you.

L1 beginner

Educating

It explains things step by step. It shows how things are connected, teaches concepts, and checks your understanding. It slows down when you're unsure.

"We do RPC here because…"
L2 intermediate

Collaborative

They work alongside you. They offer suggestions, and you make decisions together. They build on the basics and fill in the gaps where things aren't standard.

"I have three options—which one do you prefer?"
L3 advanced

Autonomous

It performs tasks and reports results. It only interrupts when decisions are made that could go wrong. Logs, not lectures.

"Done. Diff and two follow-up questions."
L4 expert

Supervisors

It runs in the background and handles everything on its own. It only alerts you in the event of a CRITICAL issue: security incident, compliance violation, or blocking risk.

"CRITICAL. I pulled the deployment."
For any field

Not just for
software development.

AISHA manages software deliveries today because that’s how it was created. But the knowledge base, expert rules, and the entire orchestration framework aren’t tied to any specific domain—you can feed it legal research, clinical protocols, marketing playbooks, research notes, or HR onboarding. It loads the structure, you set the rules, and the conductor then manages your lifecycle.

Software Law & Compliance Clinical Research Marketing Operations Finance & Accounting HR & Training Medical Clinics Architecture & Construction Custom Knowledge Base
↳ Provider pool · resolved dynamically based on complexity
OpenAI
gpt-5 / gpt-4o / gpt-4.1-mini
Anthropic
claude-opus / sonnet-4
Google
gemini-2.5-flash / 2.0
xAI
grok-3-mini
vLLM
Qwen3-30B · self-hosted
Auto-discovery
Daily scan · benchmark evaluation
Cost-cutting as a feature

Three levers.
Until −80%
on tokens.

A smaller model can handle most of your queries. Only a small percentage warrant the largest one. AISHA knows this and keeps an eye on the threshold—you pay the minimum, and the quality remains.

01 Batch
50%

Both OpenAI and Anthropic have Batch API with a 50% discount. Aisha automatically moves queries that don't require an immediate response to the batch queue. Asynchronously, transparently, at half the cost.

−50%
at OpenAI & Anthropic · async batch endpoint
02 Local
auto-diag

Aisha uses benchmarks to automatically determine which model is most suitable. Whatever the local model—Qwen3 on your GPU, Ollama, or Gemma—can handle, it sends there. It only uses a cloud model when it makes sense to do so.

$0
for questions where your self-hosted model is sufficient
03 SoulForge
+ Gateway

LLM Gateway routes traffic uniformly across all providers. SoulForge before sending condense prompt — Summarizes the history and selects only relevant KB chunks. Fewer tokens, same answer.

∼3–6×
a shorter prompt · while maintaining the accuracy of the response
Example of routing a single query · illustration

"Refactor this 800-line component for me."

1
classifyMessageComplexity()
verdict: complex · score 5.4
→
2
get_adaptive_model_tiers()
candidate: gpt-4o (best score / $)
→
3
route_task()
no compliance gate · proceed
→
4
unifiedChat() · OpenAI
$0.0184 · 1.4 s · 2,412 tokens

The same query with a static call gpt-5 would cost $0.082. Saved 4.4×. Logged to Langfuse, audit_journal, and NocoDB. You can see this in the Appsmith dashboard.

Three-tier architecture

Aishin's hands,
eyes and voice.

AISHA operates on three levels. Each has a clear role, with no overlap. You see a clear dashboard; behind the scenes, it processes the data.

01
Orchestrator

Conductor · n8n · /functions/v1 · MCP

layer
↑ DRIVING ↑
02
Dashboard & UI

Appsmith · StoryLoop — a fully customizable intranet stack

view
↑ READS / WRITES ↑
03
Data & Analytics

NocoDB · PostgreSQL · audit_journal · pgvector

data

Self-management loop

Monitoring — Health check every 5 minutes. State-change detection, no polling. 0 tokens when stable.

Self-healing — Workflow drift between n8n and Forgejo is corrected automatically. Restart, retry, circuit breaker.

Self-learning — Every day at 3:00 a.m., Langfuse analyzes traces and generates up to 3 improvement proposals. The risk router decides which ones will be auto-approved.

Model Discovery — Daily scan of API providers, automatic benchmarking, and ranking updates. The better model is used immediately.

Signaling System

As a Conductor
decides.

Each signal triggers a specific playbook. The severity level determines the degree of autonomy—ranging from a log to a STOP+escalation. It only blocks when it really makes sense to do so.

story_created low

Onboarding Playbook

get_story_context → compose_context → match_experts → suggest_next_step → generate_copilot_instructions.

pr_opened medium

Compliance Gate

Delegation to Compliance Agent → verdict. PASS: in_review. FAIL: suggested corrections. ESCALATE: human admin.

workflow_drift_detected high

Self-deployment & drift recovery

admin_n8n_workflows(compare_all) → sync_from_repo → retry once → fail again → CRITICAL escalation.

compliance_fail critical

Enforcement

Classify as critical, major, or minor. Critical = block story. Major = correction plan. Minor = warning + log.

stuck_story medium

Unblocking

get_story_context + get_delivery_timeline → suggest_next_step. >3 days: escalate to a specialist.

chat_delegation low

Aisha as the third participant

1–3 sentences. [OVERRIDE] when the LLM gives bad advice. [ESCALATION] for security-related issues. Otherwise, stay silent.

VS Code extension · aisha-dirigent

@aisha is alive
in your IDE.

The developer opens VS Code, Cursor, or Windsurf. The extension connects the workspace to the central backend, and AISHA immediately sees what you're working on—including what Copilot is currently suggesting.

@aisha · chat participant
You

/compliance Should I use .from("specialist_pricing") to adjust the rate?

Copilot

Yes, you can use supabase.from("specialist_pricing").update(...) — It works.

AISHA

OVERRIDERule RPC-Only prohibits direct .from(). Use update_specialist_pricing_audited RPC — has an audit_journal and supports RLS.

terminal · setup
$ git clone https://repo.id3a.cz/aisha/evymo-ai-orchestrator.git $ cd evymo-ai-orchestrator $ npm run setup ✓ dependencies installed ✓ PostgreSQL 17 · PostgREST · Keycloak-ready ✓ .env generated · API keys imported ✓ 55 MCP tools registered ✓ dev server → http://localhost:5173 AISHA Conductor awake · ready to conduct.
RPC-Only
No direct .from()
No console.log
→ safeError()
Well, any
→ type guard
No ts-ignore
→ @ts-expect-error
No select *
explicit cols
Copilot Watcher scans every new line before when you save. It detects code written by Copilot, not just your own.
Payment · Transparent Pricing

Four ways to,
how we bill you.

AISHA is open code under the Elastic License 2.0. We can host your project, keep an eye on it, or help you with a specific task. Or you can take the code home and run it yourself—and pay absolutely nothing for the license.

Self-hosted Free

Under your
inspection.

You clone the repo, npm run setup And it's up and running. No billing, no commitment. Open code, your own API keys, your own data.

  • Full stack on-premises — Postgres, Keycloak, n8n, Ragnarok
  • Proprietary LLM keys, proprietary data warehouse
  • Community Support · Forgejo Issues
github.com
Cloud Account Free + full AISHA trial
coming soon ☕

Shared
playground.

You need an account anyway—for example, to find a specialist here. So why charge you for it? Aisha lives in our cloud; you work in your IDE with your own Copilot / Cursor. You won't cost us anything until Aisha accesses our LLM tokens.

  • Full backend stack in the EU · free
  • LLM remains in your IDE · Aisha just gives advice from the backend
  • Full AISHA trial with our tokens · coming soon
  • Specialists' Marketplace · Guild Qualifications
  • Account deletion on request within 30 days · GDPR
Create for free →
Operation & Maintenance starting at 3,250CZK / month

We operate
on your behalf.

Monthly minimum = 1 hour of specialist time. This covers operations, monitoring, and proposals. Priority resolution of security CVEs is included in the price — Your operations must be safe under all circumstances.

  • Hosting in private EU cloud
  • Monitoring · Health Checks
  • High-priority CVE patches included
  • User Approval Prior to Maintenance
Get a quote for the project →
Specialist Guild 13,000CZK / 4-hour block

Specific
quite a bit of work.

When you need an expert—refactoring, an audit, data migration, or a new rule. LLM tokens & the expert layer included in the block price. AISHA calculates the 70/20/10 revenue split on its own.

70% specialist 20% KB 10% platform
Open the guild →
Specialist Guild · For Professionals

Become
a member of the Guild.

Qualify in the areas where you're an expert. AISHA will then send you relevant case studies, clients will see you on the marketplace, and your expert rules will generate income even without any work on your part.

Start the Qualifying Round
Free cloud account

No monthly fees. The account remains yours for as long as you want to work with us in workshops—and for another year after that.

Continuation of the training

AISHA presents new case studies, asks for your opinion, and provides exercises in the areas where you're growing.

Marketplace listing

Clients see your rating, availability, and instant booking options. Contact is handled through an audited RPC.

Passive Income from KB

Your expert rules also help other projects—you’ll receive a share of the 20% knowledge attribution every time.

Note: A Cloud account without a Guild remains active for one year after the last activity; after that, we deactivate it and delete it once the GDPR retention period has expired.

Knowledge Attribution

20% of each paid invoice goes toward the knowledge base. AISHA logs every expert rule used and normalizes the weights via log_knowledge_attribution. When your expertise is used to solve someone else's project, you earn money even without working.

get_guild_members_marketplacedirectory with prices, ratings, and instant booking
create_consultation_booking_auditedproposed → confirmed → completed
rate_specialist_audited1–5 + sub-ratings · communication / delivery / value
log_knowledge_attributionexpert rule usage → KB split
A-Z Quick Path

Train
already is driving.

A single command starts PostgreSQL, PostgREST, Keycloak, n8n, Ragnarok, and Langfuse. In just a few minutes, you’ll have your own Dirigent—locally, under your full control.

Evolution · 6 chapters

Aisha is just
the next link in the chain.

Before Aisha came into being, there was the word, the meaning, the vector, the context. Here is a brief journey—not history, but evolution.

  1. 01

    From Robot to Understanding

    Čapek named R.U.R. — a robot as the embodiment of humanity, not a machine. A hundred years later, he gave Mikolov machine option word understand — store its value as a vector. Note: understand, not understand. Reason belongs to man.

  2. 02

    Vectors and Context

    King − man + woman = queen. Meaning ceased to be a dictionary entry and became a context. Words began to converse with one another—context determines what they mean right now.

  3. 03

    Everything is information

    Text, images, music, video, email, invoices, logos. Everything can be broken down into vectors. Vektor is a common language everything that can be described—and thus everything that can be compared, searched for, and linked.

  4. 04

    IAIAI the Donkey

    A machine without context just makes noise. IAIA — Inference Anything, Imagine Anything. S context, intent, and situation Suddenly, he understands: where he is standing, where he's going, and what he's supposed to deliver there.

  5. 05

    The Master and the Chisel

    LLM is a chisel, not a sculptor. It can't create a sculpture on its own. Without a hand champion, aimlessly and without taste, she just scrubs the stone. Aisha knows which one is which in that pair.

  6. 06

    To Whom Does Aisha Owe a Debt of Gratitude?

    Just out of curiosity, who was actually responsible for bringing it to life: Čapek (word), Turing (language machine), Shannon (information), McCarthy (Lisp), Codd (relational data), Cerf & Berners-Lee (network), Engelbart (interactive computing), Spärck Jones (search), Mikolov (embeddings), Hinton, Bengio, LeCun, Schmidhuber, Hochreiter (deep networks), Vaswani (Transformer), Sutskever, Hassabis, Amodei, Altman, Karpathy (modern LLM).

Reason leads. The tool is useful.