Composio : What Slips Through When You're Shipping to 250 Integrations at Once
The Reviewer for a Codebase That Never Sits Still
How Composio uses Entelligence to catch what a fragmented, fast-moving codebase can't catch by hand
Composio builds the integration and authentication layer behind AI agents, giving LLMs a single API to connect to and take real actions across 250+ third-party apps: Gmail, GitHub, Slack, Salesforce, and more. Founded in San Francisco in 2023 by Soham Ganatra and Karan Vaidya, and backed by Lightspeed Venture Partners, Y Combinator, and Khosla Ventures, Composio now powers over 100,000 developers and 200+ startups and enterprises, including Glean.
That's a lot of surface area to keep safe. Composio's stack spans a core Python/TypeScript SDK, a runtime service, an MCP server, and a web frontend, brokering credentials and OAuth across hundreds of external providers. It's exactly the kind of sprawling, credential-adjacent integration code where a subtle bug slips past a human skim on a Friday afternoon, and where shipping fast means secrets, config placeholders, and inverted conditionals can end up in a PR before anyone catches them.
Reviewing across the sprawl, not just one repo
Most review tools work fine on a single, tidy repo. Composio's codebase isn't that. It's 13 repos, three languages, and a release cadence that doesn't slow down for review. Entelligence has reviewed 1,100 distinct PRs across that sprawl, concentrated in composio (458 PRs), mercury (370), hermes (126), and frontend (87), with the heaviest sustained activity from April to August 2025.
During that stretch, the Entelligence code reviewer left 362 inline comments across 163 PRs. Its most recent activation window on record, Feb 5 to 6 2026, ran 139 review executions with a 100% success rate, zero failures, across mercury, hermes, and integrator, averaging 41 to 121 seconds per execution depending on repo. Execution-level reliability telemetry only exists for this recent window. It wasn't instrumented during the main 2025 review period, so we can speak to reliability for the recent slice specifically, not the full 2025 run.
The catches
Each finding below is the kind of thing that's easy to miss in a fast-moving, multi-language codebase, and expensive to miss in credential-adjacent code.
Findings | |
|---|---|
A hardcoded JWT bypassing auth. | Across multiple callpage action files (mercury PR #1617), the Authorization header was set to a hardcoded JWT, letting anyone with code access impersonate privileged users. Fixed. |
A live API key committed in plaintext. | A working apikey was hardcoded directly into source in |
Hardcoded secrets in an agent tool. |
|
An inverted login check. |
|
Mock data masking a broken query. | The execute method in |
The same placeholder URL bug, caught three times. |
|
Broken down by tag: of comments flagged security, 31 were left and 5 verified fixed. Of comments flagged correctness, 88 were left and 13 verified fixed.
The numbers
1,100 PRs reviewed across 13 repos
362 inline comments left, across 163 PRs
90 of those comments verified as fixes, a 24.9% hit rate
46 distinct PRs carrying a verified fix, 28% of every PR the bot commented on
139 of 139 review executions succeeded in the most recent tracked window, zero failures
41 to 121 seconds average per execution, depending on repo
Average PR quality score rose from 4.34 in January 2025 to a stable 7.0 to 7.7 range from February 2025 onward, with the largest jump landing right after review activity ramped up.
What the data does and doesn't prove
One reverted PR shows up in the data for this org: mercury PR #12860, an "emergency bypass" fix that got reverted the next day. Entelligence has no review or comment record for that PR. It wasn't caught before merge, and it's worth saying that plainly instead of leaving it out.
There's no cross-repo learning-loop story to tell here yet either. There are 0 rows in the guidelines/learnings table for this org, so unlike some of our longer-tenured customers, there's no "learned from repo A, applied to repo B" pattern on record for Composio, at least not so far.
There are also no daily analytics rollups for this org, so the quality trend above is pulled straight from per-PR scores rather than a pre-aggregated dashboard number. And activity isn't perfectly continuous: there are real gaps in the timeline (Sept to Oct 2025, most of 2026 outside one early-February burst), so this reads as an active but intermittently-engaged account, not a fully steady-state daily user.
Closing line
Across 1,100 PRs and 13 repos spanning three languages, Entelligence caught hardcoded JWTs, live API keys, and inverted auth logic before they shipped, with zero failed executions in its most recent run.