What is Azhi Flow?

Azhi Flow is a self-hosted runtime for AI workflows. Deterministic code, agents, MCP tools, knowledge and human approvals run as nodes of one durable, governed workflow. Every run shows what the agent saw, what it was allowed to do, what it did, and what it cost, and says plainly where it cannot know.

The name comes from ஆழி (aazhi, “sea”). The logo is the letter ழ (zha) drawn as one continuous line.

The problems it is built for

Problem teams hit with agents today	How Azhi answers it	State
Agent runs die halfway and redo or lose work	Durable runs on Temporal: retries by error class, cancellation, worker_offline waits, schedules	Built
A retried agent posts or writes twice	Tool gateway with an action ledger and effect classes	Built
Nobody can say what the agent was allowed to do before it ran	The run plan (azhi plan); runs with blockers are refused	Built
Policies that look enforced but are not	Policy coverage: enforced, harness-enforced or unobservable, per executor	Built
Prompt injection from untrusted data flowing into writes	Compile-time taint rule	Built
“What did the model actually see?”	A context manifest per turn	Built
Surprise model bills	Per-turn usage, cost from declared pricing or “unavailable”, step budgets, spend limits	Built
Locked to one agent harness	Pluggable executors that declare their capabilities	Built
Locked to one model vendor	Profiles choose Anthropic, OpenAI or GitHub Copilot	Built
Human sign-off on risky steps	Approval nodes in the CLI and Mission Control	Built
Answers without sources	Datasets with hybrid retrieval and numbered citations to immutable excerpts	Built (alpha embedder is lexical-hash, not a neural model)
Untrusted workflow code on workers	Signed packages, per-publisher keys, worker trust policies	Built

Who it is for

• Teams that run the same agent work repeatedly and unattended: a weekly report, every PR review, a ticket-to-release flow. Azhi adds durability, scheduling, approvals and an audit trail.
• Platform and security owners who need to know, before a run, which guarantees actually hold.
• Engineers who already use OpenCode or the Claude Agent SDK and want those tools to run as governed steps rather than as interactive sessions.

Who it is not for

Be sure before you adopt it

• It is not an in-editor assistant. If you want interactive coding in your IDE, use that tool directly. Several of those tools can run inside Azhi as executors.
• It is alpha. Workers are supported on Linux only. macOS and native Windows local mode are tested and working, and local mode is still experimental. See Project status.
• It does not merge, close or force-push anything. By design, nothing in Azhi merges.

A workflow in a glance

# Illustrative shape only. See Write a workflow for the real schema.
nodes:
  - id: fetch        # tool: ordinary code, no model, no tokens
  - id: metrics      # script: Python (uv) or Bun
  - id: analyse      # agent: the only step that spends tokens
  - id: review       # approval: a person decides
  - id: post         # tool: Slack post, recorded in the action ledger

Read Workflow-first design for why this shape matters, and What is different for what you get that is hard to find elsewhere.