SandBase Harness

English | 中文

GitHub stars Listed on deepseek-plugin.org Release Official MCP Registry Discussions CodeQL License

AI-readable project metadata: llms.txt · installation guide

A local-first runtime for AI agents. Sessions, sandboxed tools, memory, credentials, audit trails, and a built-in Console — all running on your machine or in your own infrastructure.

SandBase Harness architecture

Why

Agent SDKs handle the model loop. Production agents need more: persistent sessions, tool governance, sandbox boundaries, credential handling, memory, auditability, and a UI for humans to inspect what happened. managed-agents is that runtime layer — not a visual workflow builder and not another model SDK.

Need	What Harness provides
Run generated code safely	Local, Docker, Kubernetes, and self-hosted worker sandboxes
Inspect long-running agents	Persistent sessions, resumable event streams, audit, and replay
Control tool access	MCP toolsets, credential vaults, permission policies, and approvals
Operate any model	OpenAI, Anthropic, MiniMax, and OpenAI-compatible providers, including DeepSeek V4
Keep infrastructure yours	Local-first SQLite and file storage with no required hosted control plane

Features

• Claude Managed Agents-style /v1 API and local Console
• SQLite metadata by default for agents, sessions, environments, credential vaults, memory stores, files, skills, and API keys — local file/skill bytes in the workspace state directory
• Resumable Server-Sent Events for session replay and debugging
• One active model provider boundary configured through Settings V2
• Sandbox backends: local process, Docker (per-session containers), Kubernetes (kubectl exec/cp), self-hosted worker queue
• MCP toolsets, permission policies, built-in tools, and skill packages
• TypeScript SDK at managed-agents/sdk
• Release gate: npm run release:check

Quick Start

Requirements: Node.js 22+, npm 10+, and a model provider API key (OpenAI, Anthropic, MiniMax, or any OpenAI-compatible endpoint). Docker is optional and only needed for Docker-backed sandboxes.

git clone --branch v0.3.8 --depth 1 https://github.com/sandbaseai/sandbase-harness.git
cd sandbase-harness
npm ci
npm run build
mkdir ../my-agents && cd ../my-agents
node ../sandbase-harness/dist/index.js init
node ../sandbase-harness/dist/index.js start

init writes a workspace into the directory you run it from: an agent, a skills folder, and config.yaml, whose provider reference is the ${OPENAI_API_KEY} environment variable. start serves the API and the Console on http://127.0.0.1:3000.

Two steps finish the setup, both on Settings > Setup at http://127.0.0.1:3000/dashboard:

1. The provider. Paste your API key into the provider form and save. The page then reports that the saved configuration is not active yet, so restart the runtime — stop it with Ctrl+C and run the start command again, or use the restart button. A saved setting only takes effect at startup. If your provider is not in the list, choose the OpenAI-compatible vendor and set its base URL.
2. The model. In the Agent models panel, set the model ID your provider actually serves — deepseek-chat for DeepSeek, for example. An agent carries its own model ID, so the gpt-4o that init writes is not valid for every provider, and a wrong ID fails the turn with model_not_found.

Send the first message from the Console: open Sessions, create a session for the agent, and type into the composer. From a terminal it is one command:

node ../sandbase-harness/dist/index.js chat agent_assistant --message "hello" --tool-approval allow

chat sends that one message and exits once the turn settles; without --message it keeps the session open and streams until you interrupt it. --tool-approval allow preauthorizes the tool calls the agent may make, which the init template otherwise parks for approval and waits for a person to answer; see CLI.

The unscoped managed-agents name on npm is not this project. Until an official scoped package is announced in this repository, install only from the tagged GitHub source release shown above. Do not run npx managed-agents or npm install managed-agents.

Try it in Codespaces

Open in GitHub Codespaces

The included development container installs dependencies and builds the runtime. When the terminal is ready, start the server on the forwarded port:

node dist/index.js start --host 0.0.0.0

Open the forwarded SandBase Harness Console port, then configure a model in Settings > Setup. Codespaces usage may be billed by GitHub; the local quick start above remains free and keeps all runtime data on your machine.

Screenshots

Console overview	Settings	API reference
overview	settings	api-ref

Use the Official SDK

The runtime answers its own /v1 API on that same port, and an official Anthropic TypeScript SDK client drives it unchanged: point the client's baseURL at the runtime, give it the runtime API key, and the quickstart in examples/official-sdk runs a whole turn — message, tool call, tool result, final reply — against it. That example is executed on every pull request by tests/conformance/official-sdk-quickstart.test.ts, so the compatibility it describes is compatibility that is tested rather than claimed.

The same surface is specified in docs/api.md, and this repository's own TypeScript SDK is documented under SDK below.

CMA compatibility

Coverage of the published Claude Managed Agents contract is declared entry by entry in src/core/capabilities/matrix.ts: of the official SDK's route surface (110 method/path pairs as of SDK 0.132.1), 90 routes are mounted and 20 refuse by name; no route is deferred. Partial and Unsupported entries always name their reason.

The table below is generated by npm run docs:compat, and a contract-honesty test fails when it drifts from the matrix.

<details> <summary><strong>Full compatibility table (generated)</strong></summary> <!-- compat-table:start -->

Area	Official capability	Status	Notes
headers	compatibility-header-admission	Partial	Version, beta, and mutual-exclusion rules are enforced for any request that carries a compatibility header. A request with no compatibility header is accepted as a local caller, which the published contract does not define; this header-free path is a deliberate local-first extension for a self-hosted single-tenant runtime, recorded as such in the headers contract §4, and it is why this entry stays partial.
headers	extension-namespace-exclusion	Supported
pagination	opaque-cursors	Partial	A collection's envelope follows the mount: the operations router serves /v1 with {data, prev_page, next_page} and its /v1/x mirror with the local {data, has_more, first_id, last_id}, chosen through one pager so no handler emits both spellings, and cursors that are readable base64url JSON rather than opaque binary. Every canonical /v1 collection serves the canonical envelope, with no exceptions: the complete-set listings carry {data, prev_page: null, next_page: null} and the windowed ones carry a followable cursor — /v1/sessions pages by number under {order, created_at bounds, page} and rejects a cursor replayed under another ordering or creation window, /v1/sessions/{id}/events carries {order, filter, after_id}, /v1/skills and the credential audit listings use {offset, filter} — so a cut page says so instead of looking complete. The one listing that is neither shape is /v1/environments/{id}/work-items, a windowed extension that adds a counts object and is named in the contract.
pagination	cursor-query-binding	Supported
errors	structured-error-envelope	Supported
agents	agent-crud	Supported
agents	model-object-profile	Partial	String and object model forms parse field by field. effort and speed are stored, returned by the read projection (the agent read, the version read, and the session snapshot), and executed on the Anthropic provider under a model capability table — effort becomes output_config.effort, fast becomes speed: "fast" with the fast-mode beta, and adaptive-thinking models receive thinking: {type: "adaptive", display: "omitted"}; a listed model refused a level or speed it cannot take fails admission, an unknown model id or non-Anthropic provider sends nothing, and a deployment's own reasoning_effort model setting is operator-level and separate. inference_geo is refused by name with unsupported_model_field because this runtime has no inference-geography control; and a canonical multiagent roster is refused by name rather than executed.
agents	multiagent-roster	Unsupported	In scope but not implemented (decisions/0002): a canonical multiagent block (multiagent_20261001 or any other shape) is refused by name on both agent create and agent update, because no thread, subagent, workflow, or advisor surface exists to honour it and the surface is a protocol of its own this runtime does not implement; accepting it would let a caller believe delegation or workflow runs are in effect. Local delegation is registered separately as an extension.
agents	local-delegation-subagent	Supported
sessions	session-lifecycle	Supported
sessions	initial-events	Supported
sessions	prompt-caching	Supported
sessions	session-update	Supported
budget	session-budget	Partial	Consumption is priced in integer microcents from the append-only log, and a session may declare a max_list_cost ceiling at creation. The builtin loop checks the ceiling inside a turn: the step that crossed the cap is the last one, the session idles with stop_reason budget_reached and a session.usage immediately before it, and an accepted budget update or removal resumes the session on its own — a tool call the ceiling stranded is settled so the resumed turn sees a paired transcript. At the ceiling the next work-starting event is refused with budget_reached while events that settle work already in flight are still accepted, so the next model request does not start; a declared outcome's revision loop reads the same spend and stops at the ceiling too, closing the outcome with result budget_reached rather than starting another grading p