For hedge funds

Coding agents for hedge funds

Your strategy code is the firm's core asset, so Nimbalyst runs Claude Code and Codex on each researcher's own machine, against your own checkout, with every agent change reviewed line by line before it lands.

Download Nimbalyst

Free for individuals. Agents run locally with your existing Claude Code or Codex subscription. Team collaboration is a paid plan.

Other downloads: Apple Silicon Mac Intel Mac Windows Linux
Nimbalyst workspace showing code, a research document, and a diagram open alongside a coding agent session

What an investment firm needs before it lets agents near the codebase

Quant researchers, portfolio managers, and the small engineering teams behind them write enormous amounts of code that never goes through a normal software process. Signal research, backtests, data pipelines, risk overlays, and execution tooling all get built quickly by people whose primary job is not software engineering. Coding agents are an obvious accelerant for exactly that kind of work.

Adoption then stalls on two objections that have nothing to do with model quality. Security and compliance will not approve uploading proprietary strategy code into a vendor's cloud sandbox, and no risk-aware desk wants unreviewed agent output landing in production research. Both come down to where the code runs and whether a person saw the diff.

Nimbalyst is built around those constraints. Agents run locally under the researcher's own credentials, the repository stays on the machine and in your network, and every change an agent makes surfaces as an inline red and green diff you accept or reject file by file. The apps are open source, so your security team can read how they work rather than take a vendor's word for it.

Why Nimbalyst

Built for firms that cannot ship their code to a sandbox

Agents run on your machine, not a vendor's

Agents run on your machine, not a vendor's

Claude Code and Codex launch locally against your own checkout, under your own subscription or API key. There is no metered cloud sandbox holding a copy of your repository, and no vendor environment your strategy code has to be uploaded into.

Every agent change reviewed before it lands

Every agent change reviewed before it lands

Inline red and green diffs, file by file, across code, research notes, configs, and data models. Accept, reject, or edit each change individually, so no agent output reaches your research codebase without a person approving it.

Permission controls per action

Permission controls per action

Decide what agents are allowed to do on their own and what requires an explicit approval prompt, from file writes to shell commands to external tools. Tighten it per project rather than trusting a single global setting.

A visual workspace for researchers who are not engineers

A visual workspace for researchers who are not engineers

Quants and PMs work in editors instead of a terminal. Markdown research memos, spreadsheets, data models, Excalidraw diagrams, and code sit in one workspace the agent already has context on.

Choose the agent, bring your own key

Choose the agent, bring your own key

Run Claude Code and OpenAI Codex side by side, with OpenCode and Copilot in alpha. Work through the provider your firm has already reviewed and contracted with rather than one a tool hardcoded for you.

Parallel research streams on one board

Parallel research streams on one board

Run several agents at once, each on its own git worktree, so a signal rewrite, a backtest harness change, and a data pipeline fix proceed in the same repository without colliding.

A record of what changed and who changed it

A record of what changed and who changed it

Full restorable history on documents, complete session transcripts for every agent run, and one-click restore to any earlier version. When someone asks what the model did in March, the transcript is there.

Shared trackers for the research pipeline

Shared trackers for the research pipeline

Hypotheses, bugs, tasks, and decisions on a shared board with your own stages. Teammates update it and so do their agents, so the board reflects the state of the research rather than last week's standup.

Shared specs your agents read and write

Shared specs your agents read and write

Real-time multiplayer markdown for strategy specs, model documentation, and decision records. Each researcher's local agent reads the same shared file, so context moves without anyone pasting it between terminals.

Comparison

Cloud agent platforms vs Nimbalyst

Feature Cloud agent platforms Nimbalyst
Where your repository lives Cloned into the vendor's sandbox On the researcher's machine, in your network
Where the agent runs Vendor-managed cloud compute, metered Locally, under the researcher's own credentials
Model provider Usually fixed by the platform Claude Code, Codex, OpenCode, or Copilot, with your own key or subscription
Review of agent output Pull request at the end of a run Inline red and green diffs, accepted or rejected per file
Non-engineer usability Terminal or developer-first UI Visual editors for markdown, spreadsheets, diagrams, and data models
Research artifacts Separate document and ticket systems Specs, diagrams, and trackers in the same workspace the agent works in
Audit trail Platform-side logs you do not control Local session transcripts plus restorable document history
Security review Vendor questionnaire Desktop and iOS source on GitHub under MIT, readable line by line
Team sync Vendor holds plaintext project data End-to-end encrypted, and only what you promote to shared
Cost model Per-seat plus metered sandbox compute Free for individual use; you bring your existing model spend

FAQ

Frequently asked questions

Does our code leave the building?
The repository stays on the researcher's machine. Nimbalyst does not clone your code into a hosted sandbox and does not keep a copy of your checkout. The coding agent still sends the prompts and code context it needs for inference to whichever model provider you have configured, such as Anthropic or OpenAI, exactly as the underlying CLI does on its own. Firms that have already reviewed and contracted with a model provider are approving the same data path they approved before.
Can our security team review the software?
Yes. The Nimbalyst desktop and iOS apps are MIT licensed with the full source on GitHub, so your team can read how the app talks to the coding agent, build it themselves, and fork it if they want changes. The collaboration server used for team sync is licensed separately. Nimbalyst is SOC 2 Type 2 certified, and the trust page covers encryption, subprocessors, and data handling in detail.
What is shared with the team, and what stays private?
Only what you promote. Documents, mockups, diagrams, and trackers you choose to share become multiplayer, and synced content is end-to-end encrypted. Agent sessions, local files, and everything you have not shared stay on your machine.
How do we stop an agent from changing something it should not?
There are two layers of control. Permission settings decide per project what an agent may do unattended and what requires an approval prompt, covering file writes, shell commands, and external tools. Every change it does make then appears as an inline diff you approve file by file, with a one-click worktree option so agent work happens on an isolated branch until you merge it.
Will quant researchers who avoid the terminal actually use it?
Nimbalyst was built for exactly that gap. Researchers work in visual editors for markdown, spreadsheets, diagrams, and data models, direct agents from a chat panel, and review changes as colored diffs rather than terminal output. Product managers adopted it for that reason, and quant researchers run into the same constraint: deep domain expertise, real code output, no appetite for a command line.
Can we keep Jira, Confluence, or an internal wiki?
Yes. Agents can reach those systems through MCP integrations, so nothing has to migrate on day one. Teams that want the tracker and the specs next to the code adopt the built-in shared trackers and documents over time instead.
Which coding agents does it support?
Claude Code and OpenAI Codex are fully supported and can run side by side, with OpenCode and Copilot available as alpha providers. Different desks can standardize on different agents without leaving the workspace.
What does it cost?
Nimbalyst is free for individual use, and you bring your own Claude or OpenAI subscription or API key for the model spend you are already paying for. Team collaboration is the paid tier, with pricing on the pricing page.
What platforms does it run on?
macOS, Windows, and Linux on the desktop, with a native iOS app for reviewing sessions and diffs away from the desk.

Learn More

Related guides and comparisons

Nimbalyst for teams

Shared docs, shared trackers, one connected workspace.

How collaboration works

What is shared, what stays local, and how it is secured.

Trust and compliance

SOC 2, encryption, subprocessors, and data handling.

Open source

The desktop and iOS apps are MIT licensed and on GitHub.

Coding agents for teams

The agent-neutral team workspace story.

vs Superconductor

Local agents and shared work vs metered cloud sandboxes.

Claude Code for product managers

The same no-terminal workspace, for non-engineers.

Parallel Claude Code sessions

Run several agents at once with a worktree each.

Visual editor for coding agents

Inline AI diffs across markdown, mockups, diagrams, and code.

Give your researchers coding agents that run locally and never land a change without review

Download Nimbalyst

Free for individuals. Agents run locally with your existing Claude Code or Codex subscription. Team collaboration is a paid plan.

Other downloads: Apple Silicon Mac Intel Mac Windows Linux