Kolega Labs 开源本地优先编码智能体 kolega-code，可自行编排多智能体工作流

Kolega Labs

ENGINEER kolegacode.com · open source

The production engine.

A local-first, provider-agnostic coding agent built for work that is too large for one context window. A planner decomposes the task, specialist sub-agents run parallel tracks, and every phase is journaled, reviewable and resumable.

# install · pypi · python 3.11+

pip install kolega-code && kolega-code .

Local-first

Your machine, your keys, your repo.

Open source

The same engine we run on client work.

Provider-agnostic

No lock-in to a single model vendor.

Journaled

Every run leaves an audit trail.

# gigacode

Spawn a whole engineering team in your terminal.

One command fans out a coordinated org of specialised sub-agents. A planner decomposes the work, independent coders take parallel tracks, and a reviewer consolidates before anything touches your branch. The run is written down as it goes.

kolega-code . · gigacode on

$ kolega-code .

/gigacode on

› migrate the loan API from Flask to FastAPI

⊢ planning agent · workflow: pipeline (3 phases)

◐ coder[routes] rewriting endpoints

◐ coder[schemas] pydantic models

◐ coder[tests] parity suite

⊢ reviewer agent · consolidating diffs

✓ phases journaled · resumable · diff ready to review

→ handing off to devsec · scan before merge

# the org it builds

plannerDecompose & sequence

Reads the codebase, splits the task into phases, decides what can run in parallel.

├─ fan out ─┤

coder

routes

coder

schemas

coder

tests

└─ converge ─┘

reviewerConsolidate & hand over

Resolves conflicts across tracks and presents one diff for a human to approve.

Why it matters on regulated work: a cross-cutting migration is not one long conversation, it is dozens of dependent changes. Fanning them out is how the work finishes; journaling them is how it gets signed off.

# control

Read-only until you say otherwise.

Plan mode investigates and proposes without touching a file. Build mode applies reviewed diffs. Shift+Tab switches between them, and anything outside the working tree (a shell command, a network call, a write beyond scope) asks first. On client engagements that boundary is not a preference, it is the control that makes the agent auditable.

scope

Permissions per session

Grant once, for one path, for one run. Nothing persists silently.

local

Your machine, your keys

Sessions, tokens and transcripts stay on the box you ran it on.

loan-servicing-api PLAN MODE

› where is interest accrual calculated?

read 41 files · no writes

accrual/daily.py:88 → ledger/post.py:214

shift+tab → build mode

⚠ permission required

run pytest tests/accrual · write accrual/daily.py

allow oncedeny

✓ diff staged · 2 files · awaiting your review

# what a run leaves behind

Every run is evidence.

Transcripts, phase journals and diffs are written to disk as the work happens. On a regulated change that record is not a nicety, it is the paper trail, and it is the same artefact set CAM™ certifies against.

.kolega/run-2026-08-14T0912/

├─ transcript.md human-readable log

├─ run.jsonl machine-readable trace

├─ phases/ plan, status, resume point

├─ diffs/ every change, per phase

└─ approvals.json who allowed what, when

How this feeds CAM™ evidence →

# no vendor lock-in

Bring the model you're allowed to use.

Frontier, hosted or fully on-premise. In regulated environments the constraint is usually where the data may go, not which model is best this quarter, so the engine treats the provider as configuration.

Frontier APIs

Hosted model APIs

Your cloud

Managed endpoints in your tenancy

On-premise

Open-weight models, air-gapped

Mixed

Cheap models for grunt work

Built for work one agent can't cover

Cross-cutting migrations, multi-service refactors, parallel feature tracks on one repository, and framework upgrades that touch every layer at once.

Where it sits

The middle of the toolchain, not the whole of it.

Coding at the speed the market expects.

We deliver with Kolega Code, and we teach your engineers to work the same way. Open core, managed when you need scale and support.