I find the Gemini Flash 3.6 a very interesting model. It is reasonably competent but most importantly super duper fast. Additionally I already pay for the Google AI Pro account so I have a nice amount of free tokens.
Unfortunately, I am a single harness kind of person, all my engineering, clawing and so on is done in my harness https://term-llm.com/
Ever since agy was released I have been keen to add it as a provider to my harness, I already do so for claude -p and cursor-agent and grok, all have their own quirks, but were reasonably straight forward to get going.
agy though has been a real tricky one. I felt extremely uneasy going with approaches I have seen out there. Ripping the OAuth keys out of the binary is wrong, parsing internal Sqlite database feels wrong.
A few weeks ago a new release of agy was made with --output-format=stream-json so I was very excited to easily wire this into my harness.
The gaps
Sadly, it is not easy.
The general pattern I (and others) have used for “binary as a endpoint” is:
- Disable the native tools that the agy harness provides
- Wire up an MCP with tools that my harness provides
Sadly there are 2 huge gaps as of release 1.1.10
- You can not disable native tools AND enable an MCP see this issue.
Native tools + MCP - works
tools: [] + inheritMcp: true - due to a bug MCP unavailable
- MCP has a non configurable knob where if an MCP tool returns 4096 bytes or more, the harness will put the result in a file, which is not desirable.
The workaround
agy vial localharness makes calls to daily-cloudcode-pa.googleapis.com term-llm adds a TLS intercepting request filter, via HTTP_PROXY. With this proxy in place I can:
- remove all native tools (so only MCP tools remain)
- undo spill over truncation which is not configurable
I now added Gemini 3.6 Flash to my rotation of LLMs I use for code reviews and quick prototyping.
Looking forward to the 2 tiny changes needed in agy that will let me remove the proxy.
