Rune Box: an environment where your agent can roam freely but with receipts.
The open-source governance harness for AI agents, by Runeward.
Put enforceable policy, human approvals, isolated execution, budgets, and signed evidence around
any AI agent. Rune Box works with an existing agent or multi-agent framework rather than requiring
a new model or orchestration stack.
Prove it in one command
Prerequisites: a running Docker, OrbStack, or Podman engine and the runebox binary.
The command creates .runebox/quickstart.toml, checks the policy, runtime, image, and state path,
runs an allowed command, proves a destructive command is denied before execution, and verifies the
signed audit trail. It never overwrites an existing policy unless --force is passed.
doctor and dashboard readiness also resolve required secret sources. A Charter that references an
unset env:// value is not presented as launch-ready.
runebox doctor quickstart
runebox --config-dir .runebox serve
runebox evidence export quickstart -o run.json
runebox evidence verify run.json
What Rune Box adds
| Concern | Container alone | Rune Box |
|---|
| Tool calls | Executes what the process requests | Checks every shell, code, file, network, and browser action first |
| Risky actions | Application-specific | allow, deny, or require-approval with an attributed decision |
| Network | Usually open unless separately configured | Deny-by-default hostname policy; strict L3 enforcement on Kubernetes |
| Limits | CPU/memory | Wall-clock, exec, egress, token, cost, and retry-loop budgets |
| Audit | Runtime logs | Append-only, hash-chained, Ed25519-signed events |
| Handoff | Ad-hoc logs and folders | Workspace tar, recovery snapshots, and portable signed evidence JSON |
| Agent identity | One opaque process | Tenant, actor, parent run, provider, model, and durable run lineage |
| Interfaces | Runtime-specific | CLI, REST, MCP, web dashboard, Kubernetes CRDs, and local SDK adapters |
Every governed action follows one path:
agent request → policy → human approval when required → limits → sandbox → signed audit event
Naming
Documentation and UI use familiar terms first. Existing API paths and file fields retain the
original themed names for compatibility.
| Plain-language term | Rune Box name | Existing surface |
|---|
| Sandbox | Citadel | /v1/citadels, Kubernetes Citadel |
| Policy file/profile | Charter | /v1/charters, *.toml profile |
| Approvals | Conclave | /v1/conclave |
| Signed audit trail | Chronicle | /v1/chronicle, [chronicle] |
| Network controls | Perimeter | /perimeter, [network] |
| Budgets and limits | Rationing | [rationing] |
| Agent group/fleet | Cohort | /v1/cohorts, [cohort] |
| Signed policy bundle | Archive | runebox archive, [policy_bundle] |
See the full naming and writing convention.
Install
Choose the package that matches how you use Rune Box:
| Install with | What it installs | Command |
|---|
| Homebrew | Rune Box CLI for macOS or Linux | brew install Runeeward/tap/runebox |
| PyPI | Python client and agent-framework adapters | python -m pip install runee-box |
| npm | TypeScript client and agent-framework tools | npm install @runebox/sdk |
For normal local use, install the CLI with Homebrew. For an agent integration, install the SDK for
its language as well. The pip and npm packages connect to a running Rune Box API; they do not
replace the CLI/runtime.
Homebrew — CLI
Local sandboxes require a running Docker, OrbStack, or Podman engine.
brew install Runeeward/tap/runebox
runebox version
runebox quickstart
pip — Python SDK
Requires Python 3.9 or newer. The base client has no third-party runtime dependencies.
python -m pip install runee-box
python -c "import runebox; print(runebox.__version__)"
npm — TypeScript SDK
Requires Node.js 18 or newer.
npm install @runebox/sdk
npm ls @runebox/sdk
See Adapters for LangChain, CrewAI, LlamaIndex, OpenAI Agents, Strands,
Vercel AI SDK, and LangChain.js installation options.
Other CLI installation options
The signed macOS/Linux installer requires
cosign so it can fail closed while
verifying the checksum manifest. Windows binaries are available from
Releases.
curl -fsSL https://raw.githubusercontent.com/Runeeward/RuneBox/main/install.sh | sh
To build the current main branch, use Rust 1.95 or newer (the workspace MSRV):
git clone https://github.com/Runeeward/RuneBox
cd RuneBox
cargo build --release --locked -p runebox-cli --bin runebox
./target/release/runebox version
cargo install --locked --path crates/runebox-cli installs it into ~/.cargo/bin instead.
Use it with an agent
Expose governed tools to an MCP-capable IDE or agent:
{
"mcpServers": {
"runebox": {
"command": "runebox",
"args": ["mcp", "--config-dir", ".runebox"]
}
}
}
Or place an agent CLI inside a sandbox and run one or many governed workers:
runebox cohort --agent claude --model sonnet build "Build a tested API"
Adapters are included for LangChain, CrewAI, LlamaIndex, OpenAI Agents, Strands, Vercel AI SDK,
and LangChain.js. See Adapters and agent groups.
The dashboard also has a read-only Live chat TTY for each Citadel. Agent harnesses publish
user, assistant, tool, and system turns with runebox_publish_conversation (or the REST,
Python, and TypeScript equivalents), and authorized teammates can follow the redacted conversation
without terminal input access. Rune Box cannot infer private UI chat text that the agent client does
not publish; wire the publish call into the harness turn callback. The publisher must connect to the
same runebox serve control plane as the dashboard (through /mcp, REST, or an SDK).
Harness agents and subagents
Rune Box is the enforcement boundary around an agent, not the component that decides how the agent
reasons. Route the tool calls of a parent agent and each delegated subagent through Rune Box to give
them explicit policy, approval, isolation, budget, and evidence boundaries.
Existing concepts keep their meaning: a Cohort is a group of peer workers sharing a task board;
it is not being renamed to “subagents.” The orchestrator still decides when to delegate, while
Rune Box records the parent/run/provider lineage and prevents a child Citadel from widening its
parent's tenant or Charter. Every participating agent can receive its own Citadel and Chronicle.
See Agent harnessing.
Policy workflow
Policies support built-in glob rules, CEL, OPA/Rego, and signed OCI bundles. Test them in CI, start
from a reviewed scaffold, or derive exact proposals from verified production evidence:
runebox policy scaffold package-approval
runebox policy test quickstart --case 'tool=shell,action=rm -rf /,expect=deny'
runebox policy learn run.json > proposed-policy.toml
policy learn never edits a policy automatically. It verifies the evidence first, skips redacted
actions, produces exact matches, and requires a human to review and broaden them.
Security posture
- The server binds to loopback by default and requires authentication before a non-loopback bind.
- Non-loopback HTTP also requires TLS unless
--allow-insecure-http explicitly acknowledges that a
trusted reverse proxy terminates TLS.
- Multi-principal RBAC scopes sandboxes, agent groups, recovery snapshots, and dashboard views to
their tenant while attributing every operation to its actor. Static tokens and OIDC JWTs use the
same authorization model, and embedded HTTP MCP shares the REST ownership checks.
- Browser automation is experimental and disabled by default. Enable it only in a trusted deployment
with
RUNEBOX_ENABLE_EXPERIMENTAL_BROWSER=1 after reviewing the security model.
Browser-capable Charters declare capabilities = ["browser"]; the dashboard then exposes governed
rendered-text and screenshot actions and their policy/egress results.
- An optional browser IDE (code-server inside the environment + ticketed reverse proxy) is similarly experimental:
RUNEBOX_ENABLE_EXPERIMENTAL_IDE=1, Charter [ide], Dockerfile.ide targets
ide / ide-agents, examples ide-demo / ide-claude / ide-codex / ide-cursor.
Limits: not per-keystroke policy; no Cursor/Claude Desktop/Codex GUIs inside the environment; no
first-class GitHub Copilot on code-server. See Browser IDE and the
security model.
- Per-action policy applies to tool calls routed through the control plane (REST, MCP, dashboard
file/shell/code actions, and SDKs). An interactive terminal or a process already running inside a
sandbox is a direct sandbox session: it receives isolation/network/resource controls and terminal
recording, but its individual commands are not intercepted for approval. Use governed tool calls
when command-level policy and signed verdicts are required.
- Report vulnerabilities privately using SECURITY.md. Rune Box remains pre-1.0;
residual limitations are tracked in ROADMAP.md.
What's inside the environment
Beyond per-action policy, a Charter can confine the agent process itself and keep its work:
- In-sandbox confinement.
runebox-supervisor runs as PID 1 in the sandbox image and launches
the agent under Landlock filesystem rules, a seccomp filter and a dropped identity, from the
Charter's [filesystem] and [process] sections. See Isolation and the
security model.
- Credential brokering.
[[endpoint]] and [[credential]] entries let the egress proxy inject
real secrets into approved requests while the agent only ever sees a placeholder, and an
endpoint's binaries restrict it to the listed executables (attributed per connection by the
in-sandbox supervisor, strict egress only). See the security model.
- Persistent workspaces and forks.
[workspace] persist = true keeps a workspace across sleep,
wake and control-plane restarts; forks start from a copy. Agent turns, Cohort memory, handoffs
and [schedule] runs build on it. See Workspaces and context.
- Hardened mode.
--hardened (or RUNEBOX_HARDENED=1) refuses to start unless every hardening
check passes, and calls that never ran return a standard refusal object. See the
CLI reference.
Architecture
Rune Box is a Rust Cargo workspace (crates/) that builds four binaries:
| Binary | Runs | Crate |
|---|
runebox | CLI, control plane (serve), MCP server, Kubernetes controller and webhook | runebox-cli |
runebox-egress | Policy-enforcing egress proxy and credential broker (Linux sidecar) | runebox-egress |
runebox-supervisor | In-sandbox PID 1: Landlock, seccomp, dropped identity (Linux) | runebox-supervisor |
runebox-browser | In-sandbox headless-browser driver (Linux) | runebox-browser |
The library crates split along the control loop:
runebox-charter: Charter schema, loader, linter and signatures.
runebox-policy: glob/argv, CEL and Rego engines, guardrails and usage accounting;
runebox-policybundle distributes signed OCI policy bundles.
runebox-controlplane: governed execution (policy, approval, limits, backend exec, ledger),
workspaces, forks, agent turns, Cohort memory and schedules.
runebox-backend: Docker, Podman and Kubernetes sandbox backends.
runebox-server (REST, WebSockets, dashboard) and runebox-mcp (stdio and streamable HTTP).
runebox-k8s: runebox.runeward.io CRDs, the Citadel/Cohort controller and the admission webhook.
runebox-ledger, runebox-keys, runebox-auditsink, runebox-termrec: the signed,
hash-chained audit trail, its keys, sinks and terminal recordings.
runebox-authz, runebox-secrets, runebox-credentials, runebox-fleet, runebox-obs:
RBAC/OIDC, secret sources, CLI login, the Cohort task board, and logs/metrics.
Release binaries ship for Linux (static musl, amd64/arm64), macOS (amd64/arm64) and Windows
(amd64, CLI only); container images are multi-architecture and cosign-signed.
Documentation
Contributions are welcome; see CONTRIBUTING.md. Licensed under
Apache 2.0.