aidashos

A local-first agent OS

Make Plans, Not Prompts

aidashos turns a design document into governed agent work on your own machine: compiled plans, isolated worktrees, your project's own test commands, cross-vendor review, and an approval gate before anything merges. Closing your laptop never loses aidashos system state.

  • Local-first
  • Durable
  • No cloud backend
  • AGPL-3.0
View Source

Public developer preview. No signup and no hosted account. The repository is the product.

Hand your local agent these prompts to install aidashos

Three prompts, in order, for any AI tool with shell access (Claude Code, Codex, or anything else). They drive the same scripts and leave the sign-ins and big-download confirmations to you.

1

Boot the stack

Checks for a checkout, then runs the boot sequence: local models, subscription sign-ins, configuration, and verification.

Read the prompt
Help me set up aidashos, a local-first agent OS.
First inspect the current workspace for an existing aidashos checkout and preserve any local changes.
If there is no checkout, ask me which parent directory to use, then clone https://github.com/rahul-nath/aidashos.git into a new aidashos directory there.
Never overwrite, delete, reset, or repurpose an existing directory; if that destination exists, ask me to select the existing checkout or choose another location.
Work from the verified checkout root.
Read scripts/boot/README.md and docs/onboarding/ONBOARDING.md before acting.

Then complete the boot sequence:

1. Run ./scripts/bootstrap.sh --check-only and report what is missing.
2. If uv or the Python environment is not ready, run `make` and wait for it to finish.
3. Run the boot stages one at a time from scripts/boot/, in numeric order (10-check-prereqs.sh, 20-install-llama-cpp.sh, 30-fetch-model-qwen3.sh, 31-fetch-model-gemma4.sh, 40-login-anthropic.sh, 41-login-chatgpt.sh, 50-set-default-stack.sh, 60-verify-boot.sh), rather than boot.sh, so each failure is visible and fixable.
4. Stages 30 and 31 download model weights and are the only slow, large steps. Before each, run `./scripts/download-models.sh --list` to show me exactly which repository and file it will fetch, and wait for my confirmation.
5. The two sign-ins are interactive browser flows. Hand control to me for them and never enter credentials yourself.
6. If a stage fails, read its output: every blocked line prints the command that fixes it. Fix, then re-run the stage; all stages are idempotent.
7. Finish by running ./scripts/boot/60-verify-boot.sh and show me its full output.
2

Start the runtime and prove it

Brings the resident services up, re-checks readiness, and compiles the example design document into a governed plan.

Read the prompt
The aidashos boot sequence finished on this machine. Bring the system up and prove it works:

1. Run ./scripts/start-agent-runtime.sh. It exits non-zero naming any service that failed to come up; report what started.
2. Run ./scripts/first-run-check.sh and show me the report.
3. Compile the example design document into a governed plan: uv run agent-ledger compile_design_doc docs/examples/work_unit_acceptance_design_doc.md. Show me the compiled milestones.
4. Run uv run pi /ledger and summarize the current state.
5. Stop there. Approving a plan and dispatching work are my decisions; tell me the exact commands (uv run pi /approve-most-recent, uv run pi /dispatch) and what each will do, then wait.
3

Attach your own AI tool

Connects Claude Code, Codex, or any MCP client to the coordination ledger, so your daily driver can operate the OS.

Read the prompt
Connect my AI coding tool to the aidashos coordination ledger over MCP.

1. Read skills/operate-agent-os/SKILL.md and docs/AGENT_MANUAL.md in this repo.
2. If my tool is Claude Code: the repo ships .mcp.json, so open a session in the repo root and confirm the agent-os server's tools are listed.
3. If my tool is Codex: show me the [mcp_servers] block from the SKILL file, wait for my approval, then add it to ~/.codex/config.toml.
4. For any other MCP client: the server command is `uv run agent-ledger --root <repo> serve`, stdio transport.
5. Verify the connection by calling read_execution_ledger, describe_resident_loops, and list_dispatch_intents, and summarize what they say about this machine.
6. From now on, prefer the ledger tools for reading system state, and drive workflows with the pi commands described in the SKILL file.

One governed path

Give the agent the task. Let the system remember the work.

Your AI drives it

Claude Code discovers the local ledger from the repository. Codex and other stdio MCP clients attach to that same record.

Use your subscriptions

Codex and Claude Code run through their installed CLIs. The default setup needs no per-token model API integration. Provider limits still apply.

Give local models real work

A local junior model handles its assigned routine judgment. Plans and approvals remain in the local ledger.

Resume from evidence

Plans, attempts, artifacts, reviews, and approvals live in local Postgres. A lost agent context does not become lost system state.

Design to decision

Four steps, one durable record

  1. 01

    Compile

    Turn a design document into an immutable, hashed plan.

  2. 02

    Implement

    Run bounded work in an isolated Git worktree.

  3. 03

    Review

    Use a separate review session to inspect the exact diff.

  4. 04

    Approve

    Stop at the operator gate before integration.

Beside your current cockpit

Teach your coding agent when to hand work over

Your interactive agent follows the operator skill. Workers dispatched by AiDashOS follow a smaller internal execution skill.

Paste into AGENTS.md, CLAUDE.md, or your equivalent project instructions

When a task needs durable state, separate implementation and review, operator approvals, recovery, or evidence that must survive this session, route it through AiDashOS and follow <AIDASHOS_ROOT>/skills/operate-agent-os/SKILL.md.

Use a direct single pass for a bounded local change.

Operator skillDispatched-worker skill

Questions with real answers

Does it run in the cloud?

The control plane runs on your machine with a local Postgres ledger. Local models run on your hardware. Configured frontier agents connect to their providers through your installed CLIs and accounts.

What happens if I close the laptop?

Plans, attempts, artifacts, reviews, and approvals are stored in Postgres. A sleeping laptop pauses local computation. When you return, the system can use its retained state to recover work; an interrupted agent may need a new attempt.

What do I need to run it?

macOS is the supported platform today. The quickstart covers the local runtime, model setup, and attaching your AI tool. Large downloads and interactive sign-ins stay under your control.

Is aidashos finished?

No. This is a public developer preview, and bugs and incomplete features remain. Your coding agent can help inspect failures and debug setup or milestones. Read the open issues or get involved.

Give the next hard task a plan.

View Source