Product
How Bahulam works.
Three product surfaces — Chat, Bahulam Code, and Workspaces — sitting on one platform built for reliability. Here's what each does and what runs underneath.
How does Bahulam work?
Bahulam routes every request through Bahulam Code sub-agents (Explorer, Planner, Coder) or a direct Chat response, depending on the surface. Each sub-agent picks the model that fits its role, runs in a sandboxed Docker or Firecracker MicroVM, and verifies its own work with tests before reporting done — self-correcting on failure instead of stopping at the first error.
Surfaces
Chat
Public AI surface.
General-purpose chat for reasoning, writing, and product thinking. When the conversation moves to code, hand the thread to Bahulam Code and keep going in your terminal.
Learn moreBahulam Code
Coding agent, in your terminal.
Bahulam Code reads your codebase, plans changes, edits files, runs tests, and self-corrects on failure. Bring your own model — 40+ supported. Self-hostable.
Learn moreWorkspaces
Ship real things, not one-offs.
Purpose-built surfaces — Twitter marketing, 3D design, BI reporting, and more. Each workspace composes coding, tooling, workflows, and chat into one place with a purpose. Not a run — a place.
Learn morePlatform
Sub-agents
Explorer reads. Planner designs. Coder writes. Each sub-agent uses the model that fits its role — a fast model to scan, a reasoning model to plan, a coding model to write. You compose the team; Bahulam Code handles the routing.
Bring your own model
40+ models supported via direct providers or OpenRouter — Anthropic, OpenAI, DeepSeek, GLM, Tencent, Google, and more. Bring your own key, pick the model that fits your budget and quality bar, switch mid-session if you change your mind.
Public benchmarks
58.7% on SWE-bench Lite. 67.8% on Verified. Every run's raw artifacts are public on GitHub, every patch is reproducible via Docker, every failure mode is named. Numbers you can verify — not marketing claims.
Sandbox execution
Agent work runs in Docker or Firecracker MicroVMs. Nothing touches your host unless you approve it. Six permission modes — from strict approval to full autonomy — let you tune the safety envelope per project.
Cross-session memory
Bahulam Code keeps a project-local .kepler/ folder with tasks, checkpoints, approvals, and durable notes. Resume a session next week and the context is still there. No prompt-stuffing to catch the agent up on what it already knows.
MCP + tool ecosystem
Model Context Protocol servers plug in as first-class tools. GitHub, Notion, Postgres, Sentry, custom services — all reachable from within an agent run without wrapping them yourself.
Self-hostable
No vendor lock-in. Deploy the runtime on your infrastructure, keep your data in your VPC, pin the CLI version. Or use the hosted platform when convenience matters more than isolation.
One balance, three surfaces
Chat, Bahulam Code, and Workspaces share the same credit pool and BYOK configuration. What you set up once works everywhere. Simple pricing that scales from a single developer to a team.
Common questions
Read the full FAQIs the whole platform self-hostable, or just Bahulam Code?
Just Bahulam Code (the coding agent runtime) today. Deploy it on your own infrastructure and keep your data in your VPC. Chat and Workspaces are currently hosted platform surfaces.
What models does Bahulam Code support?
Bahulam Code routes to 40+ models via direct providers or OpenRouter, including Anthropic, OpenAI, DeepSeek, GLM, Tencent, and Google. Sub-agents can each use a different model — a fast model to scan, a reasoning model to plan, a coding model to write.
How do Bahulam Code's sub-agents work?
Bahulam Code splits a run across sub-agents by role: Explorer reads your codebase, Planner designs the change, Coder writes it. Each uses the model that fits its role, and you compose the team — Bahulam Code handles the routing between them.
Ready to see it work?
Install the CLI in a minute. Sign in with a browser. Ship your first change today.