Openadapt Agent MCP Server

OpenAdapt Agent

mcp-name: io.github.OpenAdaptAI/openadapt-agent

OpenAdapt is a compiled program for GUI writes with no API. This package invokes it over MCP.

This package is the default runtime interface for a calling agent. The CLI remains. Headless does not mean the UI is gone.

Three roles: the calling agent operates; a named human is authority; compliance samples seals. Computer-use agents are the user of OpenAdapt. They are not the executor inside OpenAdapt.

Flow still runs the program. Policy, identity, verification, pauses, repair, and audit stay in openadapt-flow. A healthy call goes through Flow's governed run. HALTED, refused, timeout, and error come back as those outcomes. Don't summarize any of them as success. Never summarize halt as success.

Install

Claude Code / Cursor, three lines. The server generates the public synthetic tutorial at serve time. It is not in the wheel.

claude mcp add openadapt -- \
  uvx --from 'openadapt-agent[tutorial]' openadapt-agent \
  serve --allow-run

--allow-run is an explicit opt-in. Admission stays fail-closed.

The halt demo is openadapt quickstart --break-it. After the verified run, the same certified bundle hits a backend that paints a success banner and rejects the write. An independent system-of-record read catches the lie. The record did not change.

Local unsigned replay is free. If the tool returns unsigned success, treat it as failure. Production success without a Seal is failure. Don't name this skill "computer use".

python -m pip install --upgrade openadapt
openadapt quickstart
openadapt-agent serve --allow-run

Python 3.10 through 3.12. pip install openadapt now includes this package. pip install openadapt-agent is the library install if you already have a private compiled bundle. The halt demo is still openadapt quickstart --break-it.

uvx openadapt-agent serve --bundles /path/to/bundles          # read-only
uvx openadapt-agent serve --bundles /path/to/bundles --allow-run

Serve a bundle

--allow-run with no --bundles records, compiles, and certifies the synthetic MockMed workflow, then keeps that app up so a governed run can hit a live system of record. --tutorial is the same path without implying run tools. Private customer bundles still use --bundles. Those stay on the operator's disk and are never shipped in this package.

openadapt-agent serve --allow-run --runs-dir /tmp/openadapt-runs

The client gets list_workflows, get_workflow, get_run_report, list_needs_attention, and get_attention_item. Those are read-only until --allow-run. The synthetic tutorial registers run_local_quickstart.

Add --allow-run and the server registers one typed run_workflow_<opaque-id> tool per loadable bundle. Declared parameters are required. Recorded demonstration values never appear in the schema and are not reused unless you ask for that.

By default every MCP response is safe to render outside the protected workflow-data boundary. Labels, recorded values, paths, raw reports, observed text, stdout, stderr, and local exception messages stay on the OpenAdapt machine. The client gets opaque IDs, fixed outcome copy, declared parameter names and types, and count or boolean metrics.

Two development switches, kept apart on purpose:

• --allow-protected-export sends raw local metadata and evidence to the MCP client. Use it only when that client is trusted and inside the same protected data boundary.
• --allow-synthetic-recorded-defaults lets omitted parameters reuse recorded values. It requires --allow-run. Synthetic demonstrations only. Production runs require every declared parameter so a wrong-record action cannot hide in a default.

Finish a halt from the same server

Point the server at Flow's qualified deployment config if a local operator needs to finish an exception and continue the same durable run:

openadapt-agent serve \
  --bundles /opt/openadapt/bundles \
  --runs-dir /var/lib/openadapt/runs \
  --allow-run \
  --allow-attended-actions \
  --config /etc/openadapt/deployment.yaml \
  --headed

The queue summary is safe to show in an agent UI: opaque IDs, typed categories, counts, and signed-capability metadata. No observed text, no workflow values, no reports, no local paths.

Every mutation needs the opaque queue-item ID, the item's current capability digest, a caller-stable idempotency key, and an action-specific true confirmation. Before the server submits that to Flow, it opens an MCP form elicitation and the local operator has to accept it. That's a host-mediated confirmation, not cryptographic proof of who clicked. Flow records the effective local OS account as the operator.

Clients without form elicitation cannot execute attended actions through this bridge. The same Continue, Skip, Reject, Teach, and Escalate capabilities stay available through Flow's attended console:

openadapt-flow console \
  --attend \
  --allow-actions \
  --bundles /opt/openadapt/bundles \
  --runs /var/lib/openadapt/runs \
  --config /etc/openadapt/deployment.yaml \
  --headed

Flow rechecks the signed capability, run identity, bundle version, checkpoint lineage, authorization, live state, and effect evidence at decision time. Stale capabilities and uncertain delivery are refused. The same idempotency key returns the prior terminal decision instead of repeating it.

--allow-attended-actions without --config still registers Reject, Teach, and Escalate. Continue and Skip wait until Flow can construct the deployment-bound live verifier and backend. --headed is required for web Continue and Skip.

MCP tools

Run outcomes

Every run_workflow_<opaque-id> call returns one of these:

get_run_report returns a PHI-safe status and a count-only summary for a run this server created. The persisted report stays in the local operator experience unless you enabled protected export. A client must never summarize halt, refused, timeout, or error as success.

A timeout isn't a rollback. Don't retry until you've looked at the durable run.

Emit an Agent Skill

openadapt-agent emit-skill \
  /tmp/openadapt-agent-demo/bundle \
  --out ~/.claude/skills

This wraps Flow's own skill emitter, keeps its portable bundle, and adds MCP invocation, Needs Attention, and result-handling guidance. The frontmatter description is the same sentence as server.json and llms.txt. The skill is named from the workflow slug, never "computer use". If the tool returns HALTED, tell the user the record did not change.

A first-party copy lives at skills/openadapt-gui-write/SKILL.md. A folder emitted from a private bundle isn't a sanitized derivative. It includes the compiled bundle. Treat that as protected workflow data.

Trust boundary

This is a local stdio bridge. It inherits the local user's OS permissions, and that user is recorded as the operator for attended decisions. Don't expose its stdin and stdout as an unauthenticated network service.

Remote transport, account identity, tenant isolation, fleet policy, and managed execution belong to OpenAdapt Cloud. They aren't duplicated here.

Other fixed boundaries:

• Run tools and attended mutations stay off until you enable them at start.
• Target, deployment policy, timeout, and model-egress posture are fixed at server start, not supplied per MCP call.
• Parameters go through a mode-0600 temporary file, not process arguments.
• Workflow and run IDs are opaque on the default MCP surface.
• Recorded parameter values never enter tool schemas, unless you put the server in synthetic-default demo mode.
• Protected reports, names, values, paths, subprocess output, and exception text stay local unless you enable protected export for a trusted client in the same data boundary.
• Attended action schemas accept no free-text challenge answers.
• Attended mutations require protocol-native form elicitation. A boolean from an autonomous tool caller is not enough on its own.

The complete contract is in docs/DESIGN.md.

Product state

An exact Agent release enters Production only through an active signed, expiring, and revocable release admission. A missing, expired, revoked, mismatched, or unverifiable admission produces not actively admitted. The validator doesn't restore an older admission or assign a fallback lifecycle label. Check the current signed Production record.

Package history

Before v2 this repository wrapped model-driven GUI agents. That execution path now lives in openadapt-flow. The current name stays because the package bridges MCP and Agent Skills. It isn't an MCP-only package.

The public capability is the server. serve --allow-run generates the synthetic MockMed bundle at serve time; it is not vendored. A user's compiled workflow is their private artifact, supplied at launch with --bundles and never embedded in the package or a registry listing. See docs/DISTRIBUTION.md.

Machine-readable launch manifests sit at the repo root: server.json (MCP registry), manifest.json (MCPB / Smithery), llms.txt (assistants). Registry-launched installs start read-only. Execution tools wait for --allow-run.

Development

pip install -e ".[dev]"
ruff check src tests scripts
pytest -q
python -m build
python scripts/check_release_artifacts.py dist
python scripts/check_dist.py dist/*
npx -y @anthropic-ai/mcpb@2.1.2 validate manifest.json
npx -y @anthropic-ai/mcpb@2.1.2 pack . openadapt-agent.mcpb

License

MIT. See LICENSE.