Extensible harness

DeepSeek Harness

Composable sandbox-first harness with web, headless, SDK, and multi-agent workflows.

Status: Active
License: MIT
Evidence: Documented + code-verifiable, 15 sources
Product record checked: 2026-08-13

At a glance

DeepSeek AI's MIT-licensed developer-preview harness combines a local Web UI, a headless CLI, a Python SDK, multi-provider routing, durable sessions and compaction, policy-controlled repository tools, OS file-effect sandboxing, skills, optional MCP, and parallel or continuable subagents.

Good choice if

Users who want one local coding harness across a browser workspace, one-shot headless tasks, and a Python API
Teams composing model providers, tools, skills, MCP servers, approval policy, and runtime behavior as replaceable plugins
Longer repository tasks that benefit from compaction, durable sessions, background jobs, and parallel or continuable subagents

Check before choosing

DeepSeek Harness is in developer preview and explicitly warns that compatibility-breaking changes will occur
The default CLI profiles confine filesystem effects, but the sandbox does not govern network or process visibility; enforcement can be partial on older Linux kernels and Windows
Danger-full-access bypasses confinement and approvals, while the documented Python SDK example deliberately uses that posture with a bare local filesystem backend

See 3 more considerations

Out-of-tree plugins and MCP servers extend the trusted execution surface; the MCP bridge is optional and bridges tools only, with failOnStartupError disabled by default
Durable session logs, interrupted-turn recovery, resume, fork, and compaction preserve conversation state, but no product-supported project-file rollback is documented
The repository has broad engineering, end-to-end, sandbox, and provider tests, but BENCHMARK.md only explains how to run the SDK against separate tasks and supplies no admissible comparative result

Quick facts

Interfaces: Terminal, Web, Automation
Approval style: Policy rules
Model portability: Provider + local
Isolation: OS sandbox
Agent organization: Multi-agent runtime
Local model path: DocumentedThe harness does not bundle a local model runtime; endpoint availability and model capability remain external.

View full technical record

Catalog layer: Coding harness
Harness membership: Qualifies: 4 required criteria evidenced
Product role: Extensible harness
Agent organization: Multi-agent runtime
Runtime posture: Sandbox-first
Isolation options: OS sandbox
State model: Session-based
Context lifecycle: Managed context lifecycle
Permission posture: Policy rules
Verification: Tool-assisted checks
Observability: Structured traces
Recovery: Session resume
Interfaces: Terminal, Web, Automation
Provider posture: Multi-provider
Model portability: Provider + local
Subscription access: Not documented
Enterprise access: Not documented
Local model path: DocumentedThe harness does not bundle a local model runtime; endpoint availability and model capability remain external.
Claim boundary: Harness capability only

Capability support

Documented first-class product support, checked against the sources below.

External tools (MCP): OptionalOptional plugin bridge for MCP server tools over stdio or streamable HTTPSource · checked 2026-08-13The bridge is not mounted in the default profile, covers tools only, and can activate without tools when startup discovery fails.
Reusable skills: Available by defaultFilesystem-backed reusable skills in the shared default profileSource · checked 2026-08-13Skill support does not establish portability, package quality, safety, or adoption.
Local models: DocumentedCustom OpenAI-compatible self-hosted endpointsSource · checked 2026-08-13The harness does not bundle a local model runtime; endpoint availability and model capability remain external.
Agent parallelism: Available by defaultOne-shot, continuable, forked, parallel, and workflow-driven child agentsSource · checked 2026-08-13Provider choice, depth, concurrency, tool scope, and lifecycle configuration determine the effective delegation boundary.
Runs without an open UI: DocumentedNon-interactive or automation surfaceSource · checked 2026-08-13The source establishes the mechanism, not its quality or availability in every mode.
Browser control: Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.
Isolated execution: Available by defaultDefault file-effect sandbox for shipped CLI profiles on Linux, macOS, and WindowsSource · checked 2026-08-13The sandbox does not govern network or process visibility, enforcement may be partial, and danger-full-access plus custom SDK compositions can bypass it.
Undo file changes: Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.

Getting started

Run `npx @deepseek-ai/dsh@0.1.0-rc.6 web`, configure a DeepSeek, catalog, or custom OpenAI-compatible provider, choose a workspace, and review the active permission preset before starting a task.

Open official documentation

Classification and operating model

Category fit and technical mechanisms are evidence records, not product-quality scores.

Category fit: Qualifies, 4/4 criteria
Operating model: 7/7 layers documented

Inspect category criteria and operating mechanismsFirst-party records

Why it qualifies as a coding harness

This confirms category fit, not product quality. Every required criterion links back to first-party evidence.

Qualifies4 of 4 required criteria evidenced

Adaptive agent loop
Documented
The system repeatedly observes results and chooses the next action instead of following a fixed one-pass graph.
EvidenceArchitecture and agent loop
Repository tool execution
Documented
The system can use tools to inspect and change a repository or its execution environment.
EvidenceWeb UI user guide EvidenceDefault profile composition
Task-aware context management
Documented
The runtime assembles, updates, compacts, retrieves, or persists task-relevant context while work proceeds.
EvidenceCompaction subsystem EvidenceSession persistence
Model-independent runtime control
Documented
Permissions, budgets, interruption, policy, or stop controls operate outside the model's own text generation.
EvidenceDefault profile composition EvidenceProcess sandbox EvidenceModel provider configuration

How it works under the hood

Seven mechanisms mapped from first-party records. These labels describe what the harness provides, not how intelligent its model is.

7/7layers documented

Execution & isolationSandbox availableDocumented mechanism, not a performance score.
Tooling & integrationsExtensible toolsDocumented mechanism, not a performance score.
Context & stateManaged contextDocumented mechanism, not a performance score.
Lifecycle & recoverySession resumeDocumented mechanism, not a performance score.
ObservabilityStructured tracesDocumented mechanism, not a performance score.
VerificationTool-assistedDocumented mechanism, not a performance score.
Governance & permissionsPolicy controlsDocumented mechanism, not a performance score.

Measured and public context

Configuration-specific measurements and source-native activity stay separate from general product capability.

Inspect code audit, measured configurations, and ecosystem signalsContext, not a product score

Public code audit

3/5public artifacts present

Security policy: Not found
CI workflow: Present at inspected commit
Automated tests: Present at inspected commit
Evaluation assets: Not found
Contributor documentation: Present at inspected commit

The developer-preview tree contains 7,438 files, 15 GitHub Actions workflows, 1,774 test-like paths, and contributor guidance, but no repository security policy or dedicated evaluation suite. The root BENCHMARK.md only directs users to run the Python SDK against separate tasks and provides no immutable model × harness × environment × budget × attempts result, so no product score is imported.

Inspect commit 47f943859bef60e4160492346772ded9b24f765a, checked 2026-08-13

Measured configurations

No benchmark run passes the full metadata admission policy for this harness yet. Missing data is not scored as zero.

Benchmark policy and all runs

Context, not quality

Public ecosystem signals

Source-native observations for exact mapped artifacts and reviewed stable release trains. Different units and populations stay separate, and missing coverage is never treated as zero.

View this harness in Usage