Bernstein
Also known as: bernstein.run, Bernstein orchestrator, bernstein-orchestrator
Open source deterministic orchestrator that runs 40 plus CLI coding agents such as Claude Code, Codex, and Gemini CLI in parallel git worktrees with no model in the coordination loop, so runs replay byte identically behind an HMAC signed audit chain.
Bernstein is an open source governance layer for AI coding agents. It takes a software goal or a plan file, decomposes it into a task graph, and hands coordination to deterministic Python. Scheduling, worktree isolation, verification gates and merging are auditable code rather than model responses, so the same plan replays byte identically. Each task runs a CLI coding agent in its own git worktree. There are more than 40 adapters, among them Claude Code, Codex, Gemini CLI, Cursor, Aider, GitHub Copilot CLI and the OpenAI Agents SDK, so local and cloud models can mix in one run and an optional router picks a cheaper model when the task allows.
A janitor stage checks tests, types and lint, with optional cross model review and tournament selection, before any diff lands. Destructive actions pause mid run for approval in the TUI, the web dashboard, the CLI or chat platforms such as Slack, Telegram, Discord and Teams. An always on signed lineage spine, a replay journal, content addressed evidence bundles and an opt in HMAC chained audit log let a reviewer verify what happened offline without rerunning anything. Runs start from the CLI, plan files, tickets, chat commands, MCP clients, dashboard webhooks and an autofix daemon that fixes failing CI on its own pull requests.
Bernstein runs wherever Python runs, whether on a laptop, on premises, air gapped, as a Kubernetes cluster or on Cloudflare Workers. It scopes credentials per agent and supports pluggable sandboxes such as Docker, E2B, Modal and Daytona. It exposes a REST task API with an OpenAPI spec, an MCP server and an A2A agent card. The software is free under the Apache 2.0 license, built by a single maintainer, Alex Chernysh, and still in beta.
Vendor details
Canonical URL
https://bernstein.run
Category
Agent infrastructure
Subcategory
Deterministic multi-agent orchestration
Funding status
No funding disclosed; solo maintained open source project
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Bernstein has more than 40 CLI agent adapters, including Claude Code, Codex CLI, Gemini CLI, Cursor, Aider, GitHub Copilot CLI, OpenAI Agents SDK, Goose, OpenHands, and AWS Q Developer, plus a generic CLI wrapper. Sandboxes are pluggable (Docker, E2B, Modal), and artifact storage runs on S3, GCS, and R2. Chat bridges connect Slack, Discord, and Telegram. It runs in MCP server mode over stdio, HTTP, and SSE and offers a task server REST API.
In practice
A platform team points Bernstein at a refactor goal. It splits the work across Claude Code, Codex, and Aider in parallel worktrees, gates every merge on tests and lint, and leaves a clean git history.
A consultancy stands up an AI engineering crew on a client repo in minutes, keeps its own keys scoped per agent, and hands the client an HMAC signed audit record replayable for compliance review.
A reviewer checks what a multi agent run actually did by replaying its journal and verifying the signed lineage offline, without rerunning any agent or trusting a model narrative.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
10.5 / 14 capabilities · 75%
| Integrations & Tool Calling | Full |
|---|---|
|
Bernstein has more than 40 CLI agent adapters (Claude Code, Codex, Gemini CLI, Cursor, Aider, GitHub Copilot CLI, OpenAI Agents SDK, AWS Q Developer and more), plus a generic wrapper for any CLI tool, pluggable sandboxes (Docker, E2B, Modal), artifact storage on S3, GCS and R2, and chat bridges for Slack, Discord and Telegram. Bernstein docs and PyPI Aug 2026 Sourcebernstein.readthedocs.io/en/latest/adapters/adapter_guideread 2026-09-21 |
|
| Workflow Orchestration | Full |
|
A goal decomposes into a task graph, deterministic Python schedules agents in parallel git worktrees, a lifecycle FSM governs task and agent states, multi stage plans run from plan.yaml, and verified results merge automatically. Bernstein docs Aug 2026 Sourcebernstein.readthedocs.io/en/latestread 2026-09-21 |
|
| Knowledge Grounding & RAG | Not documented |
|
There is no knowledge base, retrieval or grounding layer. Agents ground in the repository itself, and the .sdd workspace holds run state. Bernstein docs Aug 2026 Sourcebernstein.readthedocs.io/en/latestread 2026-09-21 |
|
| Human Oversight & Guardrails | Full |
|
When an agent wants to take a destructive action mid-run, Bernstein pauses it for approval in the TUI, a web dashboard button, the CLI from a second terminal, or approve and reject buttons in Telegram, Slack, Discord or Teams, and always_allow rules decide which calls skip the prompt. Every merge also waits on the janitor gate, so the customer decides which actions need sign-off and routes the request into existing channels. SourceBernstein, bernstein.run/blog/operator-commands and homepageread 2026-09-21 |
|
| Security, Identity & Governance | Partial |
|
The enterprise guide documents OIDC or SAML single sign on and role based access. Per agent credential scoping gives each agent only the environment it declared, with SPIFFE workload identity and sandbox backends isolating execution. There is no SOC 2, ISO attestation or trust center. Bernstein is open source software from a single maintainer with no hosted service, and it offers its hash chained audit log in place of a SOC 2 report. Sourcedocs.bernstein.run ENTERPRISE and bernstein.runread 2026-10-25 |
|
| Observability & Auditability | Full |
|
Auditability is central to the design, with an always on signed lineage spine, a deterministic replay journal, content addressed evidence bundles, an opt in HMAC chained audit log a reviewer can verify offline without rerunning, a live TUI dashboard, and an observability extra with OpenTelemetry support. Bernstein site and docs Aug 2026 Sourcebernstein.runread 2026-09-21 |
|
| Memory & State Persistence | Partial |
|
A durable ledger resumes runs on another machine, a replay journal and content addressed evidence bundles preserve every step, and workspace state lives in .sdd with no server to provision. Learned memory across runs, in the agent sense, is not described. Bernstein site release notes Aug 2026 Sourcebernstein.runread 2026-09-21 |
|
| Deployment & Data Residency | Full |
|
Bernstein runs entirely where the user chooses. It installs via pip, pipx, uv, brew and npm, keeps state local in .sdd with no server to provision, and supports cluster mode, a Cloudflare Workers backend, Kubernetes and Docker extras, and an air gap installation profile with a wheelhouse build and deny all egress. Bernstein docs Aug 2026 Sourcebernstein.readthedocs.io/en/latest/installation/air-gapread 2026-09-21 |
|
| Prebuilt Agents, Templates & Packs | Partial |
|
Progressive skills and reusable plan.yaml multi stage plans ship with the tool, and the adapter guide covers adding agents, but there is no gallery of prebuilt agents or packaged templates. The adapters connect coding agents as workers, and Bernstein names no prebuilt roles. Bernstein docs Aug 2026 Sourcebernstein.readthedocs.io/en/latestread 2026-09-21 |
|
| Triggers & Channel Coverage | Full |
|
The web dashboard receives GitHub, Linear and Slack webhooks, a daemon installer keeps the chat bridge running after reboot, and the autofix daemon turns a failing CI run on a Bernstein-opened pull request into a fix commit, so work reaches agents from events without a person starting it. Runs also start from the CLI, from tickets, from chat commands and from MCP clients. SourceBernstein, bernstein.run/blog/operator-commands and llms.txtread 2026-09-21 |
|
| Model Flexibility & Routing | Full |
|
The customer declares each agent's adapter and model in bernstein.yaml, adapters are interchangeable across 40 plus CLI agents, local and cloud models can mix in one run, and an optional contextual bandit router picks a model and effort level per task under USD cost caps. SourceBernstein, bernstein.run llms.txt and bernstein.readthedocs.ioread 2026-09-21 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
Bernstein runs as an MCP server over stdio, HTTP and SSE, with a stateless chain anchored MCP core and an MCP Tasks surface, and offers a task server REST API with an OpenAPI reference, an adapter guide for adding custom agents and a generic CLI wrapper. Its Apache 2.0 source is itself extensible. Bernstein docs Aug 2026 Sourcebernstein.readthedocs.io/en/latest/reference/openapi-referenceread 2026-09-21 |
|
| Testing, Debugging & Optimization | Full |
|
Every agent diff passes a janitor verification stage before merge, running the repository's tests, types and lint plus an optional cross-model review, and tournament runs compare candidate outputs and select one with no model in the decision path, so only verified diffs land. SourceBernstein, bernstein.run homepage and llms.txtread 2026-09-21 |
|
| Browser & Computer Use | Not documented |
|
There is no browser or computer use capability, since the domain is CLI coding agents operating on repositories in worktrees and sandboxes. Bernstein docs Aug 2026 Sourcebernstein.readthedocs.io/en/latestread 2026-09-21 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Cooperative task cancellation in Bernstein v3.24.0 lets a canceled task keep its claim until the orchestrator sees that the agent's work has actually stopped. Immediate cancellation stays the default. The same release streams a run's journal live to the terminal UI, and the stream picks up where it left off after a server restart.
Bears on: Workflow orchestration
View sourceSign in routes in Bernstein v3.24.1 now carry a default limit of 10 attempts per client every 60 seconds, covering the dashboard, SSO sign in and callbacks, and CLI device authorization. A client that uses up its attempts is locked out, and the lockout grows longer each time.
Bears on: Security / enterprise
View sourceAdministrative runs, new in Bernstein v3.23.0, give tasks that people or outside clients claim by hand an audited container, with no planner or worker agents started. Role sessions can now be told about new tasks as they are created through a live event stream. Tasks created directly over MCP or REST also join the signed claim and completion record that orchestrated runs already use.
Bears on: Observability / auditability
View sourcePricing
Free and open source (Apache 2.0)
No billing unit. The orchestrator is free, and cost accrues in the coding agents it runs.
Included quota
No quotas or tiers. The full feature set ships in the Apache 2.0 build, including cluster mode, an air gap profile and the MCP and REST surfaces.
What is public
Everything. The tool is Apache 2.0 with public source, public docs and public benchmarks, and there is no paid tier anywhere.
Billing mechanics
No billing. Costs accrue only in the orchestrated agents (Claude Code, Codex, Gemini CLI and others) and in any paid sandbox or storage backends the user configures. Built in cost aware dispatch can enforce a dollar cap per run.
Cost watchouts
Parallel orchestration multiplies the underlying agent subscription and API costs. The sandbox backends E2B and Modal are paid third party services if used. A solo maintainer stands behind the project, with no commercial support offering.
Variable cost rationale
Bernstein itself meters nothing and costs nothing. A run's cost is the metered spend of the coding agents it fans out in parallel, through their subscriptions and API usage, and it grows with concurrency and model choice. The project exists because parallel agent bills reached hundreds of dollars a month. The control is built in. Cost aware dispatch caps spend per run in dollars. Paid sandbox or storage backends add their own charges if the user configures them.
Overage / add-ons
Not applicable. No paid product exists.
Sales call required
No, self serve available
Free / trial
Free and open source, installed with pip, pipx, uv, brew or npm
Lowest paid plan
No paid plan
Commercial notes
Open source under Apache 2.0, built by a single maintainer, Alex Chernysh, with GitHub Sponsors and OpenCollective sponsorship and no commercial support offering. The project calls itself beta software.
Key ambiguities
Nothing is open on price. Support and continuity are the open questions, since no commercial entity stands behind the project.
Support SLA / resale
No commercial support or SLA. The project is community maintained, with GitHub Sponsors and OpenCollective sponsorship.
Missing data
No commercial support, SLA, or paid escalation path exists to price
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Bernstein
The closest documented capability profiles to Bernstein among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Braintrust10.0 / 14Fuller documented coverage on Security, Identity & Governance
- Inngest10.0 / 14Fuller documented coverage on Security, Identity & Governance
- LangWatch10.0 / 14Fuller documented coverage on Security, Identity & Governance
- Thread AI11.0 / 14Adds documented Knowledge Grounding & RAG
- Windmill12.0 / 14Adds documented Browser & Computer Use
- LangSmith11.5 / 14Fuller documented coverage on Security, Identity & Governance and Memory & State Persistence
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded