A local-first AI agent harness you actually own — the loop, the memory and the eval, running on your machine.

gogo-agent is a small, readable codebase that shows the four pieces every serious agent is made of: Harness · Loop · Memory · Eval. No framework hides the parts that matter — each one is plain Python you can read, change and keep.

• Local-first. Your state lives in one SQLite file under ~/.gogo/. Open it, read it, delete it. It is yours, and nothing leaves your machine unless you connect something yourself.
• The loop is ~95 lines of plain Python: observe, reason, act, repeat. Step through it with a debugger.
• Memory is the core feature. Semantic, episodic and procedural memory, with a gate that decides whether to remember and a consolidation pass that decides what to keep.
• Watch it think. A local dashboard lights up every message as it flows through the harness, with a trace for every turn.
• Eval built in. Deterministic offline tests and LLM-as-judge runs side by side, with a release gate that blocks a tag on a regression.
• Bring your own model. Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter and more — one dialect in the loop, a small adapter per provider.

Every piece maps to a file — see the architecture and the editable diagrams in docs/whiteboards.

pip install gogo-agent
gogo                         # talk to your agent in the terminal
gogo dashboard               # …or the browser cockpit → localhost:7777

The first run tells you which key to set. To read the code and run from a checkout instead:

git clone https://github.com/DAgogo04/gogo-agent && cd gogo-agent
uv venv && uv pip install -e .          # create the env + install the `gogo` command
cp .env.example .env                    # pick a provider, paste ONE key
uv run gogo                             # talk to your agent in the terminal
uv run gogo dashboard                   # …or the browser cockpit

Try it. Say "Remember that Sam prefers morning meetings." Quit, restart, then say "Book a meeting with Sam next week." It remembers, and books the morning. The memory is the same one file from every folder: ~/.gogo/state.db.

New to it? Getting started walks the whole setup, with a check at the end of every step.

The built-in memory is local files plus SQLite — no account, no server. If you run an MCP server that publishes memory.remember / memory.search / memory.get, add it to ~/.gogo/mcp.json and gogo mirrors saved facts and research reports there too. The server is discovered by capability, so any compatible memory backend works; mark its entry "memory": true to name it explicitly.

gogo mcp                      # list servers and the account each knows you as

See integrations for remote servers, OAuth sign-in and the MCP configuration format.

How is this different from ChatGPT or Claude Desktop? Those are products you use. This is a codebase you own: the loop, the memory schema, the gate and the eval harness are all yours to read and change.

The code is MIT — see LICENSE. Bundled fonts are under the SIL Open Font License; their license files ship with them under gogo/ops/static/fonts/.