Overview
Project status, connections and the latest changes across this installation.
Activated projects
Recent activity
Setup
Register your repositories, connect their source code, and choose how AI can investigate each project.
Activate a project
Your projects
Edit project
Connect code for AI investigation
Configure a server folder or track a GitHub branch. AI inspects code when a stored answer does not fit, a question includes attachments, or someone chooses Dig deeper with AI. MCP activation alone does not grant code access.
GitHub push webhook
In GitHub, open Repository settings, then Webhooks, and add this URL with content type application/json, the secret below, and Just the push event (or select Pushes and Releases if you use release announcement rules). The service must be reachable over HTTPS. The read token does not need webhook administration permission.
Choose investigation tools
For the project selected in Source code above.
Knowledge search and attachment reading are always allowed. Code read tools are enabled once repository code access is configured. Optional tools are controlled here by the administrator and run only in a verified sandbox.
Set up optional execution tools
On the machine running Codebase Helper, start Docker with Linux containers and build the supplied tools image. Read tools work without Docker. If isolation cannot be verified, execution stays blocked.
docker build -f Dockerfile.tools -t codebase-helper-tools:1.2.0 .
Then choose Test execution sandbox. Dependencies must already be in the image; AI requests cannot install packages over the network.
Bring in existing knowledge
Agents can publish source excerpts through MCP. You can also index an existing checkout mounted on this server. Indexing reads code and documentation; it does not execute the repository or generate issue seeds.
AI checks whether stored knowledge answers your question. If it does not, investigation starts automatically. Attachments always use AI.
Your question
Answers appear here with the evidence behind them: stored sources, inspected code and any tool calls.
Browse what agents and imports have stored for a project, inspect raw records, and control which publications answers may use.
Invite Codebase Helper to a channel. It explains the one-repository rule and asks the inviter or a workspace administrator to select an activated project.
Choose projects for Slack
Only active projects appear in the Slack selector.
Turning a project off also pauses its bound-channel answers; existing bindings are preserved. Reopen the Slack dropdown to see changes.
Edit workspace connection
Use a Slack app with Socket Mode enabled. Download the manifest, create the app in Slack, and authorize its scopes. Attachment analysis requires files:read; reinstall an existing Slack app after adding this scope.
Repository bindings
A binding cannot be switched. Use a different channel for another repository. Removing and re-inviting the app preserves the binding.
Channel info command
Run /codebase-helper in a channel where the app is invited. It privately shows usage, AI status, enabled tools and that channel’s project. No AI call is made.
Existing Slack apps: download the updated manifest, add its slash command and commands scope, then reinstall the app. Socket Mode handles the command; no new public URL is needed.
Delivery queue
Save named AI configurations and choose the one used for all projects. Test a saved configuration to verify its connection and response format.
Saved configurations
Turning AI off keeps saved configurations and uses knowledge retrieval only.
New configuration
Not savedThe test checks the saved connection and structured responses. Tool calling and image analysis also depend on the chosen model. OpenAI API credentials are separate from a ChatGPT subscription.
Configure public web search and create reusable investigation tools. Publishing makes a custom tool available for project selection; each project must enable it separately.
Web fetch & search
Web fetch needs no API key and reads public HTTPS text pages. Web search uses the Brave Search API. Both are off until enabled for a project under AI permissions.
A blank field preserves the saved key. Fetch reads up to 1 MiB and returns up to 12,000 text characters; search returns up to five snippets. Pages requiring login, browser JavaScript, PDFs and private networks are unavailable. Web evidence is labeled separately from repository sources.
Custom tool catalog
New custom tool
Not savedScripts run in the protected snapshot container: /repo and the container root are read-only; /tmp is disposable. Scripts have no network, host environment or live checkout access. Use HTTP GET for APIs. Only create trusted tools: all custom tools can access saved variables. Secret echoes are redacted from results, but tools should never print, encode or return credentials.
Define values once for use in any custom tool. Changes take effect on the next investigation; no per-tool variable permissions are needed.
Environment variables
New variable
Not savedPlain values are visible here; secret values are never returned by the settings API. Both are encrypted on disk. Up to 100 variables, 8 KiB per value and 64 KiB combined. These are app-managed variables, not imports from the server’s operating-system environment. Saving or deleting a variable stops investigations using the previous settings.
The agent workflow activates a project, searches existing knowledge, publishes a meaningful change, then verifies its receipt.
Agent connection bundle
Includes the coding skill, repository instructions, a Python stdio-to-HTTP bridge, a publication example and a receipt verifier for CI.
Add the bundle to your coding agent and repository once. The skill guides publication; a required CI receipt check enforces it before merging.
Remote MCP endpoint
Create an API token
Copy this token now. It is only shown once.
Issued tokens
Publish meaningful project changes to attached Slack channels. You control the conditions, content and timing.
Delivery history
Latest 100 messages. Open an item to review its content, sources and delivery state.
GitHub release webhook setup
Choose the Releases event in GitHub under Repository settings, then Webhooks. Use JSON, this service’s public HTTPS URL and the secret below. Only published releases and subsequent note edits are used.
New publishing rule
Installation-wide controls: announcement publishing, GitHub credentials, storage and the public address.
Project update publishing
GitHub read tokens
Save a token under an alias, then choose that alias in an Auto repository. Use a fine-grained token limited to the selected repository with Contents: Read-only. Public repositories can use no token.
Saved token values are encrypted and never shown again. Reusing an alias replaces its credential for all repositories using it.
Storage
SQLite, one database per repository
Settings and Slack bindings live in a separate control database. Credentials are encrypted with an installation key. Other database adapters are not included in this release.
Download installation backupThe backup includes project data and the key needed to restore encrypted credentials. Store it securely.