For teams

Your models.
Your rules.
Your infrastructure.

Pulse is designed in Russia and for Russian teams: interface and documentation in Russian, GigaChat and local models on par with Claude Code, keys and code on your machines, policies that cannot be bypassed, and an audit that can be shown to be secure.

Three scenarios

Where do teams start?

Agents in their circuit

Foreign clouds are prohibited or questionable. The team raises a vLLM or LiteLLM gateway with qwen2.5-coder and devstral, connects GigaChat using the key - and gets the same agents with a graph, policies and swarm as on Claude Code.

  1. “Local model” → gateway address → “Check”: the server responds, function calling works.
  2. GigaChat → key from the account; CORP Pulse will select the access type itself.
  3. Swarm "Development" with GigaChat-2-Max on the backend and a local model on tests.

Review and audit

Security requires knowing what agents did with the code. Quality inspectors use a strong model to evaluate each job, the audit uploads a report on the task, politicians close what no one should.

  1. Auto-check in “Settings → Quality”: each completed work receives a rating and findings.
  2. Rules: prohibit recording outside src/, ask at git push, prohibit npm publish.
  3. Audit by task - Markdown to the security folder; in it who authorized every action.

Pilot on one repository

Take one live repository, two developers and two weeks. The goal is not to “try AI”, but to answer the question: how many tasks from the backlog do agents close to the point “they can pour in” and how much it costs.

  1. Week 1: tasks one at a time, auto mode off, budget 5 USD per task.
  2. Week 2: “With Inspection” swarm on two-part problems; Auto mode from the developers.
  3. Result: table “who writes better”, expenditure by task, audits - all from Pulse.
What does the team get?

Not “AI in the editor”, but a controlled process

Models

Any, including yours

Claude Code, Codex, Kimi, OpenCode and Grok - by subscriptions and employee keys. GigaChat - by organization key. On-premises models - via your vLLM or enterprise keyed gateway. A new model can be connected on the day of release using the “Model manually” field.

Rules

Policies by roles and tools

What is prohibited is prohibited in any mode. Reviewers and inspectors cannot govern by role definition. The budget for the task stops both the swarm and the inspectors. The rules are stored in the settings and transferred as a file.

Transparency

Graph, tape, audit

Every turn and call on the graph with rewind. Activity feed with filters by type and task. Audit report on the task - who launched what, who allowed it, which files were affected, how much was spent.

Quality

Which agent writes better - with numbers

Inspectors accumulate assessments in the “agent + model” context. A month later, the team has a table, but not impressions: who should be given the backend, who should not be allowed to access security, what adjustment in the instructions helped.

Knowledge

Repository memory

Decisions and agreements - “migrations only through alembic”, “this module is being rewritten” - are accumulated in a memory file and are sent to each agent in the first turn. A new team member, human or agent, is not starting from scratch.

Process

Swarm as a team standard

The composition of the swarm is your process, recorded in the settings: what roles, on what models, with what instructions. The instructions are being versioned. The same composition gives the same order of work on any task.

Shared board · 0.30.4

Address board: the card has a person in charge and an inspector

They are appointed by the owner or participant of the project, the observer only reads. The assignments have their own version: if two people change them at the same time, the second one receives the refusal “The assignments have already been changed. Update the board" rather than overwriting someone else's selection. The card has a discussion - notes, a request to transfer someone else's task and a request for a review with one button; assignment, reviewer role, and open requests appear in the “Requires Attention” dashboard. Receiving a card does not trigger someone else's agent: the board reports the state, and work continues in your clone after the branch is transferred. Assignments and discussion live on the Pulse Cloud server - if it is an old version, the team page will say so directly, and the board and post will continue to work.

Collaboration in documentation →

Without network

Actions are piling up in the queue

Without a connection, the board is shown from the last confirmed photo with a timestamp, rather than being blank. Appointments and appointments made offline go to “Awaiting Sending” and leave in the same order when connected. A server failure—a version conflict or no rights—leaves the action in a queue with an error text until you decide to “Remove”: other people’s changes are not silently overwritten.

Pilot

Two weeks, one repository

We personally participate in the pilot: we help connect models, configure policies to suit your rules and read the first reports of inspectors.

  1. Talk for an hour

    What models are acceptable, what cannot be released, who is responsible for the review, what repository we take.

  2. Set up in a day

    Assembly on participants' machines, connecting a gateway or GigaChat, policy rules, swarm composition, budgets.

  3. Two weeks of backlog tasks

    Real ones, not educational ones. We are in touch, comments are included in the next version - we release a patch for each delivery.

  4. Summary with numbers

    How many tasks reached the point of “can be added”, how much did it cost, which of the agents did the best job, what had to be banned.

Honestly about today

What doesn't exist yet

  • Central politicians. The rules live in the settings of each machine; are carried by the file, but are not imposed by the server. The roadmap contains a signed policy file that cannot be weakened locally.
  • Total command memory. Repository memory - on the machine; synchronization between participants via git or cloud is planned.
  • SSO and corporate login. A Pulse account is created separately; there is no company single sign-on. The application itself on the machine is single-user: the roles live in a common project - owner, participant, observer, and for the card there is a responsible and reviewer.
  • General account of the organization. Licenses are issued one per account; there is no single screen with payments and a list of employees. The contribution of participants and the result of published tasks are visible on the general project board, and the consumption is a separate setting.

If your pilot blocks any of this, please say so in your application. We release a patch for each delivery and plan the work according to what the teams request. The full list is in road map.

Application

Tell us how it works for you

We will respond within a few days and offer a time to talk. If you don’t need a pilot and just want to try it, write too, we’ll sort it out.

  • The pilot is free, you pay your providers for the models
  • We help with connecting the gateway and GigaChat personally
  • Your comments will be included in future versions.
  • No contracts before you see the result

We respond within a few days. We do not send mailings and do not share the address with third parties.

Application accepted We will respond to the specified email.