Keep the agents running. See the one that needs you.

A fast, persistent terminal workspace for Codex, Claude Code, OpenCode, and the shells around them.

_{Real Nagi build. Deterministic demo data. No mock interface.}

Nagi is a native terminal multiplexer built for parallel coding agents. Your panes keep running when the terminal disappears. Reattach locally, over SSH, or from a phone. Press Ctrl+B, then G, to see what needs you and jump straight to it.

Stay running	See attention	Automate it
Persistent panes, tabs, and workspaces	Blocked, working, idle, and done at a glance	Typed CLI and Unix socket API
Detach and reattach over SSH	Search and filter every live agent pane	Durable mission journal and worktree claims
One Rust binary, no Electron	Mouse and keyboard navigation	Headless server for scripts and agents

Why Nagi

Starting agents is easy. Keeping track of them is not.

Nagi keeps real terminal processes, their current state, and the workspace they own on one calm surface. It does not replace Codex, Claude Code, or OpenCode. It gives them somewhere reliable to run.

Works today

Surface	Status
Persistent terminal sessions, splits, tabs, mouse, SSH reattach	working
Agent cockpit with search, state filters, counts, and direct switching	working
CLI and Unix socket API	working
Guided mission creation, recovery, journal replay, and worktree claims	working
Managed Codex, Claude Code, OpenCode, and local ACP launch	working
Fresh proof execution, portable evidence packs, and guarded closure	working
Project recipe setup/check/cleanup, port leases, health checks, restart adoption, exact cleanup, and mission wiring	working
Digest-bound provider handoff with same-mission continuation	working
Nagi Dawn/Night, theme files, and Ghostty theme import	working
Native plugins	working; explicit unrestricted trust required before enablement
Plugin v2 validation, grants, immutable locks, and WASI runtime	working for zero-capability components; requested host capabilities stay denied until approved
Public plugin registry	not available yet
Signed binaries and public release channel	not available yet

The end-to-end mission loop is usable on the current development branch. Nagi is still early access because the full plugin capability broker, distribution, and the public release security gates are not finished.

Agent compatibility

Agent	Runs in Nagi	Managed mission path	Tested version
Codex	yes	guided local launch, read-only socket launch	0.144.5
Claude Code	yes	guided local launch, read-only socket launch	2.1.212
OpenCode	yes	guided local launch, read-only socket launch	1.18.3
Any local ACP agent	yes	guided local launch after explicit write consent	contract v1

Any terminal program can run in a pane. Managed launches add a mission contract, an isolated worktree, durable attention, and proof. The public socket cannot approve workspace writes or answer permission prompts: those decisions stay with the human in the local TUI. nagi doctor reports an installed provider as ready only when its version matches the tested managed runtime above. Other versions still run as ordinary terminal processes, but are not advertised as a supported managed mission path.

For Codex and Claude Code, nagi doctor --probe-providers also performs the real initialize handshake and exits before creating a thread or sending a user turn. It verifies the installed protocol without spending a model turn.

Claude Code, supervised from prompt to proof

Choose Claude Code without changing the workflow. Nagi binds the run to a mission, keeps write consent local, and carries the same mission across a provider handoff without transferring hidden reasoning.

Start with Claude. One outcome, explicit checks, isolated worktree.	Hand off to Claude. Same mission, digest-bound snapshot, fresh proof required.

Claude can finish the work. Nagi proves it finished.
_{Rendered by the real Ratatui UI from deterministic state. No browser mock.}

Build from source

The mission runtime currently targets Unix systems. You need Rust 1.96.1 and Zig 0.15.2.

git clone https://github.com/Cod-Hash-Studios/nagi.git
cd nagi

zig version  # 0.15.2
cargo build --release --locked
./target/release/nagi

Nagi uses its own config, runtime paths, sockets, logs, and environment variables. It does not reuse an existing Herdr session.

Start a mission

Check the machine, launch Nagi, then open the cockpit:

nagi doctor --probe-providers
nagi

Press Ctrl+B, then G, then n. The wizard asks for one outcome, explicit acceptance criteria, a proof command, and a provider. Write access requires a local confirmation and launches the provider inside an isolated Git worktree. When .nagi/project.toml exists, the same confirmation screen lists its setup, services, cleanup, and explicit file-copy scope before anything executes.

Run the cockpit demo

Build Nagi, make sure jq is installed, then start an isolated development server:

cargo run -- server

In another terminal:

scripts/seed_navigator_demo.sh
cargo run

Open the cockpit with Ctrl+B, then G. Use b, w, i, d, and a to filter by state. The script inserts simulated status data into real Nagi workspaces and panes, features only Codex and Claude as coding providers, and refuses to touch the main socket unless explicitly allowed.

API

nagi api schema
nagi api snapshot
nagi mission list
nagi mission get <mission-id>
nagi mission proof <mission-id>
nagi mission handoff <mission-id> --to claude-code --preview
nagi mission handoff <mission-id> --to claude-code --start --artifact-sha256 <sha256> --generated-at-millis <timestamp>
nagi mission close <mission-id>
nagi project detect . --json
nagi project validate .
nagi project setup . --yes
nagi project check . --yes
nagi project services start . --mission <mission-id> --run <run-id> --yes
nagi project resources preview

The generated mission schema lives at docs/next/api/nagi-api.schema.json. Closing a mission always reruns its declared checks. Stored output alone is not accepted as fresh proof.

Make it yours

Start with Nagi Dawn or Nagi Night, load a theme file, or import colors from Ghostty. Plugins can add actions, hooks, panes, and link handlers. Today those plugins are native processes with resource limits and a scrubbed environment. Nagi keeps them disabled until you explicitly grant unrestricted native trust; legacy registry entries migrate to untrusted. Manifest v2 plugins run as WASI components with bounded memory, fuel, output and wall time. They receive no inherited filesystem, network or host environment access. Grants are versioned, checksum-bound and revocable; capabilities without a host binding remain unavailable even after approval.

Architecture in 20 seconds

terminal clients
      │
      ▼
single-writer Nagi server
      ├── panes, tabs, sessions, render streams
      ├── mission journal, worktree claims, attention
      └── managed provider adapters
             ├── Codex
             ├── Claude Code
             ├── OpenCode
             └── local ACP agent

The TUI is a client of the server. Mission truth stays in the durable runtime, so SSH clients, headless automation, and future interfaces share one contract.

Contributing

Read AGENTS.md before changing runtime contracts and CONTRIBUTING.md before opening a pull request.

cargo fmt --check
cargo test --locked -- --test-threads=1
python3 -m unittest scripts.test_brand_isolation scripts.test_fork_safety
bun test src/integration/assets/nagi-agent-state.test.ts

Provenance

Nagi is an independent derivative of Herdr, starting from v0.7.4 at 50aaa2ec046ee26ff407c20f49de496f522512a8. Required copyright and attribution notices remain intact. See FORK.md.

Nagi is licensed under AGPL-3.0-or-later. The separate commercial license offered upstream is not granted by this repository.

Less tab hunting. More finished work.
_凪