Navi Agent

English | 简体中文

Navi Agent is a compact, self-evolving agent runtime inspired by Hermes. It focuses on a single reliable execution pipeline across runtime, tools, memory, telemetry, and evolution, with WeChat as the first gateway.

Documentation

Run Navi Agent in 2 minutes

Configuration, tools, permissions, sessions, memory, architecture, and the Weixin gateway.

Quickstart

uv tool install navi-agent
navi-agent init
# Add your model credentials to ~/.navi-agent/config.yaml
navi-agent doctor
navi-agent

Start the WeChat gateway:

navi-agent doctor --doctor-gateway weixin
navi-agent gateway start

Navi Agent stores config and runtime data under ~/.navi-agent. Use NAVI_PROFILE=work for an isolated profile or NAVI_HOME=/custom/path for a custom location.

Allow access to additional directories explicitly:

navi-agent --add-dir ../shared --add-dir /path/to/repo

WeChat

Configure ~/.navi-agent/config.yaml:

gateway:
  weixin:
    token: your-token
    account_id: your-account-id
    base_url: https://ilinkai.weixin.qq.com
    dm_policy: pairing
    allowed_users: []

dm_policy supports open, pairing, allowlist, and disabled. Manage pairing requests with:

navi-agent --gateway-pairings weixin
navi-agent --approve-gateway-pairing 123456

Evaluation

Inspect evaluations use the production Navi runtime and write replayable results under ~/.navi-agent/evals/inspect/.

uv sync --extra eval
uv run navi-agent eval run general-qa
uv run navi-agent eval run human-eval
uv run navi-agent eval run bfcl
uv run navi-agent eval run agentbench-os
uv run navi-agent eval run swe-bench-verified --limit 1
uv sync --extra terminal-bench
NAVI_EVAL_SANDBOX=modal uv run navi-agent eval run terminal-bench-2-1 --limit 1

See evals/README.md for suites and usage.

Development

uv sync
uv run pytest
uv run navi-agent