Team agents
Write a specialist once, such as a code reviewer, a test writer or a security auditor, and give it to everyone or to chosen teams. Members call it with @ in the editor and the CLI.
Understand deeply.
Build deliberately.
One agentic engineer across your editor and terminal: it understands the code you have, plans and builds the change, runs the tests and reviews the result.
From understanding a codebase to shipping in parallel.
Stay with the problem.
All the way through.
UnderstandThe whole picture.
The whole picture.
Before the first edit.
A codebase is more than the file in front of you. CodeGraph connects symbols, callers and dependencies so you can explore the impact of a change.
- Navigate unfamiliar and legacy code
- Trace callers, imports and tests before you edit
- Keep what it learns in project memory
Code Graph panel, illustrative
Callers · 3
signIn()src/auth/login.ts:42
postLogin()src/api/handlers.ts:118
loginCommand()src/cli/auth.ts:27
createSession()src/auth/session.ts:12
session.test.tscovers createSession()
Calls · 2
insertSession()src/db/sessions.ts:31
signToken()src/crypto/token.ts:8
Impact radius: 3 callers, 2 calls, 1 testHandoff to Plan
BuildFrom an intention
From an intention
to an implementation.
Plan the change, work across files, run the terminal and debug what comes back. Use a focused agent, or let Orchestrator split the work and hand it to subagents.
- Orchestrator, Plan, Code, Debug and Ask, plus your team’s own
- Multi-file edits, terminal and browser, in a sandbox if you like
- Allow, Ask or Deny for every tool it reaches for
Build, illustrativePlanReading checkout.ts
You asked
An empty cart costs nothing.
Add a test to prove it.
checkout.ts
10import { sum } from './money';1112function checkout(items: Item[]) {13 if (!items.length) return 0;14 return sum(items);15}
checkout.test.ts
4 expect(checkout(cart)).toBe(1290);5});67it('totals 0 when empty', () => {8 expect(checkout([])).toBe(0);9});
npm testSandbox on✓ totals 0 when emptyTests: 42 passed, 42 total
In parallelSeveral tasks at once.
Several tasks at once.
One branch each.
The Agent Manager gives every task its own worktree, a separate folder and branch, so agents work side by side without touching the same files. When a check fails or a reviewer comments, send it back: the agent commits and pushes the fix.
- A worktree and a branch for every task
- Pull request status beside each branch
- Failing checks and review comments sent straight back
Agent Manager, illustrative
storefrontThree worktrees · your checkout untouchedNew worktree
main
fix/empty-cartcheckout.tsCode
feat/csv-exportexport.tsOrchestrator
docs/api-errorserrors.md@docs-writer
Sent back to CodeNew in SyphaCodeMore of the work.
More of the work.
None of the guesswork.
Agents your team writes once, a rule for every tool, a sandbox for commands, memory that carries over, a record of every run and a browser of its own.
Permissions for every tool
For each tool, choose Allow, Ask or Deny. Add exceptions for paths and commands: let git run freely, never read .env files, block destructive deletes.
Sandboxed commands
Run the agent’s shell commands inside an operating-system sandbox. They can write only inside the project and reach only the network addresses you allow.
Project memory
The agent saves lasting facts about your project as it works, and compacts long sessions before they run out of room, so each new task starts with what the last one learned.
Agent runs
Follow a prompt through every subagent it started. If a run fails, see where it failed, why, and what to do next, in plain language.
Web search and a browser
Web search works with models from every provider, and the agent can open a browser to click through a page and take screenshots, under the same permissions.
Built around the way you workThe small things
The small things
that add up daily.
Context, tools, models and safeguards that make an agentic engineer useful on the second day, and the hundredth.
Context that follows the task
Reference files, folders, problems, terminal output and git changes with @ mentions. Project rules and AGENTS.md carry your conventions into every session.
CodeGraph, not text search
A live map of every symbol, caller, import chain and test. See the impact radius of a change before it merges, and hand the graph to Plan.
Skills worth repeating
Package repeatable work as slash commands: /review, /test-gen, /doc-gen and /security-audit ship with the product, and teams share a vetted library.
Tools inside your perimeter
Native MCP connects issue trackers, internal APIs and databases. Terminal and browser actions run within the permissions your organisation sets.
Your choice of models
Let Auto Mode pick from models hosted in India, or choose Claude, GPT, Gemini and 400+ more through their providers’ APIs, your own keys or local models. Organisation settings decide who sees what.
Three surfaces, one account
VS Code, the JetBrains family and the CLI share one login, the same agents and the same settings. Changes your organisation saves reach all three on their own.
Model and tool availability depends on your plan, provider and organisation policy.
Existing systems. New possibilities.Keep the knowledge.
Keep the knowledge.
Move the code forward.
Recover the business rules hidden in a legacy system. Capture its behaviour in tests, then work through upgrades and refactors in changes your team can review.
Existing code
function renewal(account) {
if (account.active) {
return account.plan;
}
return null;
}Recovered understanding
The rules behind the code.
- Active accounts keep their plan.
- Inactive accounts return no renewal.
- Document the dependencies.
Illustrative engineering workflow
SyphaCode in your terminalSame intelligence.
Same intelligence.
A different prompt.
Plan a change, edit across files and continue a session from the command line, with your team’s agents and settings. Use non-interactive commands in scripts and CI. Sypha CLI is part of SyphaCode.
Get the CLI~/your-next-project
❯
syphaReady. What would you like to build?
Sypha CLI, illustrativeQuestions, answeredA little more
A little more
clarity.
What is SyphaCode?
SyphaCode is SYPHA’s agentic coding tool: AI agents across a VS Code extension, the JetBrains family and a terminal CLI, on one account with one set of models and policies. They understand existing code, plan and build changes, run tests and review the result, several tasks at once if you like, and you approve what lands.
ProductDo I need to change editors?
No. Install the VS Code extension or the JetBrains plugin (IntelliJ IDEA, PyCharm, WebStorm, GoLand, PhpStorm, Rider, CLion and more), or use the CLI on macOS, Windows or Linux. All three share your Sypha account.
ProductWhich languages does it support?
SyphaCode works with the languages your models handle. Structural features such as codebase indexing and CodeGraph rely on Tree-sitter parsers, so widely used languages get the richest context.
ProductCan it help with existing and legacy code?
Yes. Use Ask and Plan to recover the business rules and specifications hidden in an undocumented system, CodeGraph to map dependencies, Code to capture current behaviour in characterisation tests, then modernise module by module in changes your team reviews.
ProductHow is this different from autocomplete?
Autocomplete is included, but SyphaCode is an agent. It reads across files, plans, edits, runs commands and checks the results, then shows you a diff to accept, refine or roll back.
ProductIs the CLI a separate product?
No. Sypha CLI is SyphaCode in your terminal, with the same agents, models and organisation settings. It supports saved sessions and runs non-interactively for scripts and CI pipelines.
ProductCan SyphaCode work on several tasks at once?
Yes. The Agent Manager gives each task its own Git worktree, a separate folder and branch, so agents work side by side without editing the same files and your own checkout stays as it is. Pull request status shows beside each branch, and a failing check or a review comment can go straight back to the agent, which commits and pushes the fix.
ProductWhich agents come with SyphaCode?
Five you can pick: Orchestrator coordinates work across agents, Plan plans the change before anything is edited, Code writes and edits code, Debug diagnoses and fixes issues, and Ask answers questions without changing code. Two more, Explore and General, take the searches and side tasks the others hand off. In the editor they appear as modes in the picker, and each can have its own model.
Agents and contextCan we create our own agents?
Yes. Organisation admins publish team agents, such as a code reviewer, a test writer or a security auditor, to everyone or to chosen teams, each with its own instructions, tool access and model. Members call them with @ next to the built-in agents, and an enforced agent cannot be replaced. Members can also create their own, import agent files or take them from the marketplace.
Agents and contextHow do I give it the right context?
Type @ to reference files, folders, workspace problems, terminal output, git changes or a URL. Project rules and AGENTS.md carry your conventions, codebase indexing adds relevant repository context, project memory keeps durable facts between sessions, and Enhance Prompt sharpens a vague request before work begins.
Agents and contextWhat does CodeGraph add?
CodeGraph maps every symbol, caller, import chain and test, so you can see the impact radius of a change before it merges. Hand a graph selection to Plan for an architecture review. CodeGraph is available on Teams and Enterprise.
Agents and contextWhat happens when a conversation gets long?
Auto compaction summarises older context before it reaches the model’s limit, with a model you choose, and can prune old tool output. Project memory keeps the durable facts for the next session, and you can search, resume or import past sessions.
Agents and contextWhich models can I use?
In Sypha Cloud, two kinds: open-weight models we host in India, which Auto Mode can choose between for each request, and more than 400 models such as Claude, GPT and Gemini, which we do not host but call through their providers’ own APIs via the Sypha gateway. On any plan you can also bring your own provider keys or connect local models through Ollama, LM Studio or an OpenAI-compatible endpoint, and an organisation can add its own endpoints for every member. Organisation settings decide which ones appear.
Models and keysWhat is Auto Mode?
Auto Mode lets Sypha pick the model for each request from models hosted in India; while it is on, requests go through Sypha. It is available on the Basic and Teams plans where Sypha offers it, and administrators can switch it off for a team, a role or a single member.
Models and keysCan I bring my own keys?
Yes, on free and paid plans. With your own keys, requests go directly from your machine to your provider under that provider’s terms.
Models and keysCan I use an existing Claude Code or Codex subscription?
In connected deployments, supported integrations can sign in to those accounts where the provider permits it. Those are cloud services, so they are not available in an air-gapped installation.
Models and keysCan I run models locally?
Yes. Connect Ollama, LM Studio or any OpenAI-compatible endpoint. Small models struggle with tool use, so a capable coding model gives the best results.
Models and keysCan administrators restrict models?
Yes. Provider lists apply to the organisation, each team, each role and each member, and every level can only narrow the one above. Anything off the list never appears in the picker, and endpoints the organisation adds reach members on their own.
Models and keysCan I review changes before they are applied?
Yes. Every write is presented as a diff in the editor, inline or side by side. Whether an action needs manual approval depends on your permissions and your organisation’s policy.
Review and safetyCan I undo the work?
Checkpoints return repository files to an earlier task state, and the CLI keeps undo and redo for a session. Restoring files does not reverse external actions already taken by tools.
Review and safetyWhat can the agent touch?
Each tool, from reading and editing files to shell commands, web access and MCP, is set to Allow, Ask or Deny, with exceptions for particular paths and commands, such as letting git run freely or never reading .env files. With the sandbox on, commands can write only inside the project and reach only the network destinations you allow.
Review and safetyWhat do the safety checks do?
Checks in the editor stop prompts that try to override the agent’s instructions, seek help with abuse or pull large parts of the codebase out, and file writes that contain malicious code. Organisation admins and the member’s Team Manager see an alert in the dashboard. Only a short excerpt is kept, with secrets removed, and none at all for organisations with encryption at rest. Alerts are kept for 90 days.
Review and safetyHow do secrets stay out of prompts?
Tool rules can deny reads of environment files and other sensitive paths for everyone, and safety-check excerpts have passwords, keys and tokens removed. Enterprise deployments can also keep key stores and credential files out of the local index, so they never enter a prompt or an embedding.
Review and safetyCan a developer switch the controls off locally?
No. A section the organisation or a team locks is read-only in the member’s editor and CLI, and every level below can only make settings stricter. Members’ apps pick up saved changes on their own, and if a policy cannot be fetched, access is denied rather than granted.
Review and safetyHow does automated PR review work?
Connect GitHub, opt in the repositories you want reviewed, and choose a model and a review style: Strict, Balanced or Agile. SyphaCode posts contextual comments on new or updated pull requests. Review generation uses developer credits.
Review and safety