Saha

A local work agent that receives events, runs bounded AI workers, and keeps a durable record of decisions and deliverables. Install it for your own accounts and workflow. No personal workspace IDs, credentials, conversation history or model subscriptions are included.

Status: early local pilot, version 0.2.0. Saha can run work without a prompt for every step, within configured responsibilities. It is not a complete replacement for a person, a production deployment agent, or a multi-user hosted service.

Get started

Requires Python 3.11+ on macOS or Linux. The demo needs no AI account.

git clone https://github.com/suhaib-ghl/saha-agent.git
cd saha-agent
python3 -m venv .venv
source .venv/bin/activate
python -m pip install .
saha init
saha doctor
saha serve

In another terminal, activate that venv and run saha open. Submit a demo job, answer its decision, inspect the generated note, and mark it reviewed. Then configure an authenticated Codex or Claude Code installation and select a real worker.

Your config lives at ~/.saha/config.json; your ledger, private logs and task files live under ~/.saha/state/. No config or credentials are inherited from the author. SAHA_HOME, SAHA_CONFIG, SAHA_STATE, --config and --state support other profiles. Each person must use their own credentials and permissions.

Full installation, background startup, upgrade and uninstall

What it does

Capability	Implemented behavior
Durable work queue	SQLite inbox, deduplication, scoped responsibilities and one active job
Codex	Local implementation and sandboxed commands/tests
Claude Code	Restricted file-based analysis/editing; no Bash in this adapter
Team	Codex engineer, separate Claude reviewer, at most one repair round
Desktop	Official Computer Use via Codex app-server; genuine live owner approvals
Slack	Optional Socket Mode event intake with channel and sender allowlists
GitHub / ClickUp	Explicit read-only API adapters, identity checks and change-driven enqueueing
Decisions and memory	Durable owner feedback, scoped saved instructions, reviewable artifacts
Interoperability	Read-only MCP, skills, commands and a session-status hook

The runtime is adapter-based; it currently supports Codex and Claude Code, not arbitrary providers automatically. Each CLI uses its own installed login and billing. Native desktop support is verified on macOS with a compatible official host; it is optional and not bundled. New sessions may need approval. No always-available computer access is promised.

How work starts

flowchart LR
    S[Slack push / signed hooks / local request / API changes] --> I[Verify, filter and deduplicate]
    I --> Q[Durable queue]
    Q --> W[Bounded worker or team]
    W --> A[Artifacts and execution receipts]
    W --> D[Owner decision when needed]
    D --> Q
    A --> R[Review and complete]

Slack Socket Mode and accepted webhooks wake the queue directly. GitHub/ClickUp API scanners currently poll every 30 seconds; unchanged scans make no model calls. A new installation establishes a baseline rather than replaying all old work. Local queue checks do not scan apps or invoke models. Computer-use access alone is not an event listener. Jobs still wait when another job is active, limits are reached, or the machine is asleep/offline.

A task decision starts a new worker with a durable handoff. A live tool permission answers the waiting tool in the same worker. The dashboard distinguishes these. The process of reviewing artifacts and marking complete remains explicit.

Scope and limits

One OS user owns an instance. Do not expose its dashboard publicly or share owner keys.
Jobs run in isolated directories, not automatically in company repositories.
No external messages, pushes, merges, deployments, refunds, or automatic ticket closure.
No automatic Facebook intake, native voice calls, or broad app monitoring.
Saved owner feedback changes future context; it does not retrain model weights or turn third-party messages into permission.
Default limits: 10 minutes per job, 20 job starts per UTC day, bounded attempts, 8 MB output per child and a Claude per-invocation cost limit. Codex has no hard dollar/token cap in this project. Team jobs can contain multiple model invocations.
Codex app-server is experimental. The bundled adapter must be revalidated when the vendor protocol, plugins or OS permissions change.

Documentation

Development

python -m pip install -e .
python -m unittest discover -s tests -v
python -m compileall -q saha_agent scripts tests
node --check saha_agent/web/app.js  # Node is optional for this syntax check
python scripts/package.py        # allowlisted source export, without runtime data

CI exercises the portable core on macOS and Linux. Live vendor logins, external accounts and native app permissions are not supplied to CI. See verification notes for what a passing core suite does and does not prove.

License

MIT. Third-party CLIs, hosts and services remain subject to their own terms. Installing Saha does not grant access to another person's accounts.