codehere
An open-source, local-first audit and cost layer for the AI coding agents a team already uses: what each agent cost, a signed record of what it did, and a safety hook that blocks dangerous commands before they run.
In one sentenceStarted as an AI coding agent with an audit layer; now a vendor-neutral layer across Claude Code, Codex, Aider, Gemini and OpenCode. One npm install, no account. v1.0 on npm with 1,371 automated tests. Early traction: 9 sign-ups and an inbound request for a sandbox setup.
Stack
The call I'd own
CLI over IDE extension. Owning the audit layer beats competing on raw coding quality, and a CLI is what an enterprise security review can read in one pass. IDE extension stays on the roadmap once the governance primitive is stable.
The bet
Enterprises are replacing junior engineers with AI coding agents, but nobody is shipping the apprenticeship-upgrade path. Meanwhile, DeepMind/MIT (2025) measured 17.2× error amplification in multi-agent systems, and 45% of AI-generated code contains vulnerabilities.
codehere's wedge is not "another coding agent." It's the pre-execution audit layer that every agent will need. We own the governance step, the scanner that runs before an AI write touches disk, and bundle it with a coding agent so developers get the value without an extra tool.
What it does
Tracks sessions across five coding agents and ends the day with one report: what each agent did, what it cost, and a trust score.
For Claude Code, a pre-execution hook checks every Bash, Edit and Write call in under 5 ms, blocks destructive, credential-leak and injection patterns, and appends each decision to a local audit log.
Shared project memory, so switching agents does not mean starting cold. Everything stays on the user's machine.
PM decisions I'm proud of
Changed direction once the market moved. OpenAI and Anthropic now sell enterprise agent platforms, so codehere stopped competing as a coding agent and became the vendor-neutral record for individuals and small teams that run several agents.
Scoped down hard. The original PRD was a full agentic IDE. I cut it to CLI + audit layer because owning one verifiable step beats a half-working competitor to GitHub Copilot.
Wrote a 20-scenario futures analysis (World A–D) to pressure-test positioning: if hyperscalers ship first-party audit, where do we win? If they don't, where do we win? Answer: the "bring-your-own-agent" adapter surface where enterprises already have tool proliferation.
Chose Ollama + local SQLite over managed vector DBs. Cost goes to zero on the user side, and enterprises get a story for their security team on day one.
Tradeoffs I'd revisit
Pattern-based scanning has a false-positive ceiling. Next iteration uses an LLM-judge over flagged diffs, but that re-introduces a dependency I initially wanted to avoid.
Multi-provider is a feature for early users and a tax on me. Each provider has different streaming, tool-calling, and error shapes. A thin adapter abstraction paid for itself by v0.3 but slowed v0.1.
Want to talk about codehere?
Currently taking conversations about senior AI PM and founding PM roles in the UK, EU, the Gulf, and Southeast Asia. Remote also works. Fastest reply is email.