CodeGPT

Build and run AI coding agents with CodeGPT, managing agents and loading documents into their knowledge base. Agents review, explain, and generate code grounded in your own repositories and docs.

17 actions API key available

CodeGPT website ↗ Platform Documentation ↗ Capability data verified 2026-08-01

A merge lands on the documentation repository

Agent reads the changed files and their paths from the commit

Upload Document loads each changed file into the knowledge base

List Documents compares the indexed set against what the agent holds

Agent works out which documents the replacement call would drop

Owner receives the removal list with each document's chunk count

Owner confirms the list before Set Agent Documents overwrites it

What This Integration Enables

A coding agent is worth exactly as much as what it has read. Ask a general model about your authentication layer and it will answer confidently about somebody else's authentication layer. CodeGPT's premise is that the agent should be built around a knowledge base you control, and this connector gives FlowRunner flows the whole of that premise. An agent can be created with its own system prompt, model and retrieval depth. Documents load into the workspace knowledge base, the right subset attaches to the right agent, and the conversation happens from there.

That makes CodeGPT a natural component in the Agent Factory rather than a chat window bolted onto a repository. A flow can provision one agent per team, per customer or per service, each grounded in its own documents and each reachable from wherever the question is actually asked. It can keep those documents in step with the source of truth automatically, because Upload Document takes a public URL and CodeGPT chunks and indexes the content on its side. Chat With Agent always sends the reply complete rather than as a stream, which is the right shape for a flow step: the agent answers, the flow keeps going. These operations are built and verified against the vendor's official API.

Without FlowRunner

Answers drawn from stale copies The agent explains a module the way it worked several releases ago

No visibility into what it read Nobody can list the files behind a given answer

Knowledge base updated by hand Someone remembers to re-upload docs when they remember

With FlowRunner

Knowledge base tracks the repository A merge starts the sync rather than a reminder in someone's calendar

The reading list is inspectable List Documents returns exactly what the agent can retrieve from, with chunk counts

Attachment changes are reviewed The set of documents an agent holds changes only after someone approves the diff

Use Case Scenarios

Keeping the agent's knowledge in step with the repository

A merge lands on the documentation repository in GitHub. The flow reads the changed files, calls Upload Document for each one so CodeGPT re-indexes the content, then calls List Documents to see the resulting workspace state. Before it attaches anything, it computes the difference between the document set the agent holds today and the set the sync would install, because Set Agent Documents replaces rather than appends. The removals go to the documentation owner for confirmation. Once confirmed, Set Agent Documents installs the new list and the agent starts answering from current material.

The code question answered where it was asked

A developer asks in a Slack channel why a particular retry policy exists. The message starts a flow that calls Chat With Agent against the workspace agent grounded in the service's design docs and README files, and replies in the thread with the answer plus the document titles it retrieved from. The value is not that a model answered. The value is that the answer came from the team's own written decisions, in the channel where the question was asked, without the developer opening a second tool or the original author being interrupted for the third time this month.

An agent per customer, provisioned by a flow

A new implementation project is approved. The flow calls Create Agent with a system prompt tuned to that customer's configuration, calls Extract Document Metadata on their onboarding notes to get a suggested title, summary and keywords, then Upload Document to index them, and Set Agent Documents to attach them. When the project closes, Delete Agent removes the agent while leaving the documents available to others. Update Agent can publish an agent so any user can reach it and set the PIN required to open it, which is where a second pair of eyes belongs: publishing changes who can talk to a customer-grounded agent, and that is a decision rather than a configuration detail.

Human-in-Loop Highlight

Set Agent Documents replaces the entire attachment list. Pass eight document IDs and the agent holds eight documents, whatever it held a second earlier; pass an empty list and it holds none. That is a reasonable API, and it is also the one operation here that fails without failing. A sync job that crawled half a repository because a path changed will hand over a short list, the call will return success, and the agent will keep answering every question in the same confident register with half its context gone. No error is raised, no alert fires, and the people relying on the answers have no way to tell. So the agent computes the difference first and stops before writing. It sends the knowledge base owner the documents that are about to be detached, each with the token and chunk count from List Documents, and asks: "This sync would remove 14 of the 31 documents attached to the code review agent, including the payments service design notes. Apply the replacement, or hold and re-run the crawl?" Human-in-the-loop belongs here specifically because the alternative is not a broken agent, it is a quietly worse one.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

17 actions

Conversations

Chat With Agent Sends a conversation to a CodeGPT agent and returns its reply in a single response. Streaming is disabled explicitly on every call, so the answer arrives complete rather than as server-sent events. The agent's own prompt, model, temperature and attached documents govern how it answers.

Agent Management

List Agents Lists every agent in the workspace with its ID, name, description, type, prompt, model, temperature and top-k setting. Use it to discover the agent ID needed by Chat With Agent or to audit how each agent is configured.
List Marketplace Favorites Lists the CodeGPT marketplace agents the account has favorited, returned in the same shape as the workspace's own agents. These agents can be used with Chat With Agent without being recreated in the workspace.
Get Agent Retrieves a single agent's full configuration, including its prompt, model, temperature, top-k and welcome message. Turn on Include Documents to also return the list of document IDs attached to the agent's knowledge base.
Create Agent Creates a new agent in the workspace with its own system prompt, model, temperature and retrieval depth. Only the name is required; every other setting falls back to the workspace default. Attach documents afterwards with Set Agent Documents to give the agent a knowledge base.
Update Agent Updates an existing agent, changing only the fields supplied. Beyond the creation settings this can also publish the agent so any user can reach it and set the PIN code required to open a shared agent. Returns the agent's updated configuration.
Set Agent Documents Replaces the set of documents attached to an agent's knowledge base with the supplied list of document IDs. This is a full replacement rather than an append, so include every document the agent should keep. Pass an empty list to detach all documents.
Update Agent Image Replaces the avatar shown for an agent in chat. The image is fetched from a public URL and forwarded to CodeGPT, which accepts JPEG, PNG and WebP files up to 5 MB.
Delete Agent Permanently removes an agent from the workspace. Documents attached to the agent are not deleted and remain available to other agents. This cannot be undone.

Knowledge Base

List Documents Lists every document loaded into the workspace knowledge base, with its ID, name, file type, extracted metadata, token count and chunk count. Use it to find the document IDs required by Set Agent Documents.
Upload Document Loads a file from a public URL into the workspace knowledge base so agents can retrieve from it. CodeGPT chunks and indexes the content and returns the new document ID, which can then be attached to an agent with Set Agent Documents.
Get Document Retrieves a single document from the knowledge base including its metadata, token count and chunk count. Use it to confirm a file was indexed correctly before attaching it to an agent.
Update Document Metadata Overwrites the metadata stored against a document. Title, summary, keywords and language are all required by the API; description is optional. Good metadata improves how reliably an agent retrieves the right document.
Extract Document Metadata Sends a plain text file to CodeGPT and gets back suggested title, description, summary, keywords and language without storing the file. Only .txt files are supported. Useful for previewing metadata before uploading, or for enriching a document after Upload Document.
Delete Document Permanently removes a document from the knowledge base. Agents that referenced it lose access to its content immediately. This cannot be undone.

Account

Get User Returns the CodeGPT user behind the API key, including username, full name, avatar and the organizations the user belongs to. Use it to discover the organization IDs that can be set as the service's Organization ID.
Update User Updates the profile of the user behind the API key, changing only the fields supplied. Both the username and the display name can be changed independently.

Frequently Asked Questions

What can FlowRunner do with CodeGPT?

FlowRunner agents can run Chat With Agent, List Agents, and List Marketplace Favorites in CodeGPT, plus 14 more actions.

Does connecting CodeGPT to FlowRunner require OAuth?

No. CodeGPT connects to FlowRunner with an API key, no OAuth flow required.

Can CodeGPT trigger a FlowRunner workflow automatically?

CodeGPT doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with CodeGPT

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us