Modu

A lot of teams can’t hit public registries from dev/CI Modu now supports private registries, pull-through mirrors and shared caching so agent tool installs are: fast (cache HIT), reliable (mirror HIT), and compliant (no public egress) Simple & secure as God intended

We just released a nice DX upgrade to our tool gateway: versioned tools & MCP servers. Phased rollouts and easy rollbacks With Modu everything is versioned and ships like a microservice across all of the tools your org uses. Simple & secure as God intended

We released some updates to our hardened endpoint between your team's coding agents and all your tools/MCPs. OIDC in, scoped policies + full audit out, etc. Tool-use should be safe within your org (and in general) Secure and simple as God intended

We just added better observability/auditing to our sandbox gateway Standardized p95 latency, success rate, error types per project across all of the runtimes/isolation boundaries your team uses (cloudflare, modal, e2b, daytona, etc.)

Today in Modu: we just updated our API to run your AI workloads across your favorite sandbox providers to have hard budget and safety guardrails (time, cost, concurrency, network, etc.) We think a lot about run-time security so we're rapidly iterating on scoping and more!

RT @brexton: Env vars/secrets should never be visible to coding agents, only accessible through tools. Secret abstraction is now in Modu /…

New in Modu: tool bundles, one config for all the coding agents your team uses Per repo you pick exactly which tools/MCPs, sandboxes and CLIs an agent can call and nothing else Secure & simple as God intended

We just improved the BYOK flow for all your coding agents (cursor, amp, etc.) in Modu’s package manager With the new threat surfaces that AI dev introduces, nothing should touch your repo before it passes through a global security layer, locally or in cloud containers/etc.

New in Modu: env var policies for your coding agents Per-agent, per-repo secret scopes + automatic log redaction Simple and secure as God intended

We updated our Slack integration so you can run tasks across as many agents (amp, codex, claude code, etc.) as you want in parallel with your team with proper provenance & automated PRs It's made me lazy and now I just pick the successful one-shot (this time it was @FactoryAI)

Yes and this is why gateways like @askModuAI exist

Sandboxes or agent runtime (ie e2b, daytona, vercel, modal, cloudflare) will be the fastest growing piece of infra due to agents the next few years, and we’re only at the tip of the iceberg of what all the possible needs for agents will be: dev environments, ephemeral, gpu serverless, etc Excited to share something that will make it much easier for developers (and agents) to leverage sandboxes next week!

Common flow we see: run a task across multiple agents (factory, amp, etc.) & pick what to merge The UX to review agentic work will be crucial as agents are used more but we also want to be as lightweight as possible since great interfaces will quickly evolve (ie cursor, github)

Side by side benchmarking for coding agents is here ❤️

.@askModuAI

Still a WIP before we release our real world use coding agent benchmarks broadly but notable things from the past few weeks: 1. Flippening starting to happen with OAI Codex and Claude Code since gpt5-codex released 2. Sonnet 4.5 plus major releases like new Codex/Factory/Amp/Cursor CLI, Devin rebuild, etc need to settle — expecting better performance across the board soon (UX matters a lot!) 3. Opinionated harnesses like @AmpCode @FactoryAI @DevinAI are continually increasing in merged PR % rates and starting to pass first-party labs in developer preferences for outputs/diffs to merge 4. Major cost changes like @FactoryAI’s and Gemini CLI’s generous token limits are underutilized but are super economical 5. Merged PR rates esp for one-shots on “small” tasks (sub 10k total tokens/10 min tasks) are increasing faster across the board 6. Sleeper hit that we’re excited to add and analyze: @julesagent!

OpenAI's efforts to catch up with Anthropic's code-writing AI seem to be working: OpenAI's Codex has pulled ahead of Anthropic's Claude Code assistant by some measures, and its popularity with developers is catching up too, based on new data from Modu: theinformation.com/articles/opena…

A few days into this week, even after @FactoryAI and @AnthropicAI big launches, @AmpCode still has the highest merged PR success rates (it's my personal daily driver) My prediction for the next two weeks is that Factory climbs up with CLI support and same with CC as 4.5 gets settled, but @OpenAI dev day might surprise us! Plus who knows where @julesagent will sit here too once it's added

I'm giving it a week before @claudeai Code flips @OpenAI Codex again with 4.5 in merge PR success rates Really good decision making, plus CC is non-trivially better than Codex when it comes to core tool-use (like raising PRs)

Side to side benchmarking for a given task among coding agents is the most common action for @askModuAI Releasing public benchmarks soon! Please DM if you use amp/codex/claude code/cursor/devin/etc & have strong opinions on patchviewer/diff UX, turns out it’s a heated debate 😊