OpenHands

MCP support ships on every plan as a client, with the agent connecting out to external MCP servers and using their tools automatically, and simplified MCP server controls with disabled skills kept out of agent context; documented integrations cover GitHub, GitLab and Bitbucket for source control, Jira and Linear for issue tracking and planning, Slack for messaging and PagerDuty as a workflow trigger, and the agent runs terminal commands, installs dependencies and browses the web inside its sandbox.

Sourceopenhands.dev/blog/introducing-agent-canvas and docs.openhands.devread 2026-08-30

Agent Canvas runs parallel agent sessions in one workspace and orchestrates multiple agent backends, running the built in OpenHands agent or third party harnesses including Claude Code and Codex through the Agent Client Protocol; the agent plans a task then writes and edits files, runs terminal commands, installs dependencies, browses the web, runs tests and opens pull requests, iterating until done, with sub agent delegation for specialized agents in parallel and automations decomposing issues into tasks.

Sourceopenhands.dev/blog/introducing-agent-canvas and docs.openhands.dev/sdkread 2026-08-30

The agent reads and navigates the codebase directly inside its sandbox, locating relevant code from a plain language task or a GitHub issue, browses the web for external information, and reaches connected data sources through MCP servers; repositories can be given clearer roles in Agent Canvas and Persistent Agent Memory carries context forward. No codebase indexing, embedding or retrieval architecture is documented.

Sourceopenhands.dev/blog/introducing-agent-canvas and docs.openhands.devread 2026-08-30

A confirmation policy decides when an action needs the user's approval before it runs: AlwaysConfirm gates every action, ConfirmRisky gates only actions a security analyzer rates risky, and NeverConfirm runs unattended; a pending action leaves the agent waiting for confirmation, and the user can reject it with feedback so the agent tries another approach.

Configurable security policies sit beside it, every agent runs in an isolated sandbox, and work lands as a pull request for normal review. Scheduled automations run unattended by design, which is a policy choice rather than the absence of a gate.

Sourcedocs.openhands.dev/sdk/guides/security and /sdk/arch/securityread 2026-09-29

The access surface is documented: Enterprise lists SAML and SSO, RBAC and fine-grained access control through the Agent Control Plane, SaaS or self-hosted VPC deployment, cost guardrails and centralized management, and the SDK adds a security analyzer and configurable security policies, with every agent in an isolated sandbox. A Vanta trust center is linked from the site, and the enterprise, pricing and security pages neither name an attestation nor say one is pending; no SOC 2 or ISO report is published on those pages. The trust center sits at app.vanta.com.

Sourceapp.vanta.com/openhands.dev/trust/an0kgh9goh7xyqkhsb6tlpread 2026-09-29

Built-in OpenTelemetry tracing records agent execution steps and each tool call and execution, exportable to any OTLP compatible backend the customer runs (Laminar, MLflow, Honeycomb, Jaeger, Datadog, New Relic) with transcript, tree and timeline views of prompts, tool calls, outputs and nested agent activity; a metrics guide tracks tokens, cost and latency. OpenHands Enterprise sends conversation traces to the customer's own OTLP platform, deploys Laminar for trace analysis, and lists complete audit logs for every agent action.

Sourcedocs.openhands.dev/sdk/guides/observability, /enterprise/integrations/observability-platforms and openhands.dev/enterpriseread 2026-09-29

Persistent Memory is an opt-in two tier store the agent maintains itself: a user tier under ~/.openhands/memory for knowledge across projects and a project tier under the workspace's.openhands/memory, each holding a curated MEMORY.md index injected at session start and daily logs read on demand.

The agent is instructed to record durable facts, merge duplicates and prune stale entries; the combined index is capped at about 6,000 characters with the oldest lines dropped first; and the files are plain Markdown that people can review, edit or delete at any time. AGENTS.md holds instructions rather than memory.

Sourcedocs.openhands.dev/sdk/guides/persistent-memoryread 2026-09-29

Agent Canvas is a frontend that connects to multiple agent backends, running locally on a laptop, in Docker containers, on remote VMs, on Kubernetes clusters, in OpenHands Cloud or in an enterprise virtual private cloud, with dockerless install options and documented remote backends via Modal and secure internet exposure via ngrok; the platform is MIT licensed so the whole stack can be self hosted, forked and modified, and deployment guides exist for AMD Instinct GPUs and NVIDIA accelerated compute.

Sourceopenhands.dev/blog/introducing-agent-canvas and docs.openhands.dev self hostingread 2026-08-30

Pre-built automations ship with their own documentation page and are the recommended third step of onboarding, including a Slack channel monitor and a GitHub PR review assistant, alongside documented workflow patterns for weekly repo summaries, stale issue review, dependency checks, PR summaries, test generation, docs drift checks and migration sweeps; a skills system is documented with disabled skills kept out of agent context.

Sourcedocs.openhands.dev agent-canvas prebuilt-automations and openhands.dev/blog/introducing-agent-canvasread 2026-08-30

Automations trigger from real engineering events including Slack messages, GitHub pull requests and issues, Linear tickets, Jira and PagerDuty, or run on a schedule for recurring work such as weekly repo summaries, morning stale issue reviews and Monday dependency checks; agents can also be started from the Agent Canvas console, and ngrok exposure enables event driven triggers wherever agents run.

Sourceopenhands.dev/blog/introducing-agent-canvas and docs.openhands.dev agent-canvasread 2026-08-30

OpenHands is model agnostic by design, connecting to any major provider including Claude, GPT and Gemini or to local models, with LLM Profiles letting developers switch models mid conversation and configure automations to use lower cost models for straightforward tasks; the vendor's stated position is that not every scheduled workflow should default to the most expensive model, and documented partnerships cover NVIDIA Nemotron and open models on AMD compute with vLLM.

Sourceopenhands.dev/blog/introducing-agent-canvasread 2026-08-30

A Python SDK exposes the same engine powering the CLI and hosted cloud, supporting custom toolsets including MCP and Python native tools, multi agent delegation, message passing and configurable safety and security profiles, with agents connectable to remote workspaces running locally, in Docker, on VMs or on Kubernetes; the whole platform is MIT licensed and forkable, the Agent Server can be deployed as a remote backend, and the Agent Client Protocol lets third party harnesses plug in.

Sourcedocs.openhands.dev/sdk and openhands.dev/blog/introducing-agent-canvasread 2026-08-30

OpenHands publishes a documented Evaluation Harness guide covering how to integrate a customer's own evaluation benchmark into the framework, including a configuration file pattern for defining multiple LLMs to evaluate against one another, a documented command line entry point and run_controller function, numbered steps for creating an evaluation workflow, and a recommendation to adapt the closest existing benchmark from the evaluation/benchmarks/ directory; a separate benchmarks repository provides standardized evaluation pipelines for OpenHands agents with swebench-infer and swebench-eval commands accepting custom dataset and output arguments, parallel execution in isolated containers and an optional remote runtime, currently being migrated from V0 to the V1 Agent SDK. The vendor separately publishes the OpenHands Index scoring model and SDK combinations across five categories of software engineering work, which is its own benchmarking rather than a harness the customer runs.

Sourcedocs.openhands.dev/openhands/usage/developers/evaluation-harness and github.com/OpenHands/benchmarksread 2026-08-30

The agent browses the web as a first class capability within its sandboxed environment, using a browser to gather information and verify work alongside file editing and terminal execution; the agent is described as doing the work a developer does including browsing, and this runs inside an isolated Docker container so it can act without risking the host.

Sourceopenhands.dev and openhands.dev/blog/introducing-agent-canvasread 2026-08-30