One place to see what your agents can access, what they remember, and how well their skills work.
Agent Estate is a local dashboard for tools, MCP servers, live agent context, skill evaluations, memory and memory updates. It links evaluation results to the records stored in Langfuse when that integration is configured.
Inspect context usage and its breakdown across instructions, tools, skills, memory and conversation.
Review individual checks, spot failures and inspect the evidence behind evaluation results.
Both screenshots show the real interface with synthetic demonstration data, not live sessions or evaluation results.
The dashboard includes its skill evaluator. A small synthetic example runs locally without model credentials. Model quality tests use your configured model; Langfuse is optional. Missing Herdr is explained in the interface.
Give your coding agent this instruction from the checkout:
Read
skills/agent-estate-setup/SKILL.mdand set up Agent Estate locally. Verify the dashboard, explain which connections are available, and walk me through any missing setup.
The setup skill ships with the source. It can also be installed using your agent's skill installer, selecting that folder from this repository. See Langfuse setup and verification for how evaluation records reach your own project.
Use Python 3.11 or newer. Create a local Python environment:
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python server.pyOpen http://127.0.0.1:8767. The first-run connection panel finds standard agent locations and lets you add custom skills folders, memory folders, tool folders and MCP configuration files. Save once to refresh the inventory; no shared skill index is required. The release includes built browser assets; Node is needed only for rebuilding and browser tests. See evaluation setup for adding your own skills and model tests.
See installation, integrations, local data and access, and contributing.
- Tools and MCP servers: configuration inventory, local tool-list probes and source evidence. Configured does not mean available in every live session.
- Live context: Herdr session discovery and supported harness inspection, with estimates labelled and optional captured request evidence.
- Skills: inventory, evaluation results, comparisons and explicit user-triggered runs through the connected evaluator.
- Memory: Claude, Codex, Hermes and Antigravity stores, with file and graph views.
- Updates: recent file creations and latest edits from those four stores. This is not a prose summary or full revision history.
- Traces: exact result links and export evidence from the connected evaluator and Langfuse. Independent validation of the judge is a separate requirement.
npm ci --prefix web
npm run build --prefix web
.venv/bin/python scripts/test_portable.py
npm test --prefix webCore tests use empty temporary home directories and synthetic data. Browser tests start their own local server, visit every page and verify disabled evaluation actions. The default test command does not require other desktop apps. Install a Playwright Chromium browser if prompted with cd web && npx playwright install chromium.
Existing evaluator tests remain available to developers with that integration: run AGENT_ESTATE_LIVE_TESTS=1 .venv/bin/python -m unittest discover -s scripts -p 'test_*.py'. These exercise the populated local setup and should not be confused with the portable core suite. npm run test:live --prefix web verifies an already-running Agent Estate server; set AE_URL if it is not on port 8767.
scripts/export_release.py DESTINATION prepares an allowlisted copy without Git history, local state notes, old screenshots, private results, credentials or legacy automation. The source is MIT licensed; third-party packages retain their own licenses. Review the exact resulting tree and scan it before publishing. Nothing in this script pushes or publishes it.