MCP
COBACK speaks the Model Context Protocol, so Claude (and any MCP client) can work your workspace directly. The endpoint is stateless streamable HTTP; one command connects it:
claude mcp add --transport http coback https://api.coback.ai/api/v1/mcp \
--header "Authorization: Bearer $COBACK_API_KEY"
The dashboard shows this exact command, filled in, when you create a key under Settings. With the
CLI installed, coback mcp prints it too.
Tools
Read tools, on every key:
| Tool | Does |
|---|---|
| search_brain | Semantic search over the company brain; the answer to "how does the company handle X" |
| list_cards / get_card | Enumerate knowledge cards; fetch one with full content |
| get_skills_file | The entire brain as one markdown file |
| search_corpus | Semantic search over the uploaded documents |
| list_laws | Recently ingested laws, filterable by jurisdiction and date |
| list_matches | Screening results with score, rationale, and verdict |
| list_tasks | The task board, filterable by status |
| list_products / get_product | Product compliance state and findings |
Write tools, only on read-and-write keys (a read key does not even see them):
| Tool | Does |
|---|---|
| submit_match_feedback | Record a relevance verdict on a screening result |
| create_task / update_task_status | Put work on the board, move it |
| add_comment | Comment on a task or match, attributed to the key |
Notes
- The endpoint keeps no session: every POST is one JSON-RPC message, answered as JSON.
- Tool output is markdown text, so any client can render it without a schema.
- Trust labels travel with the content: brain cards state whether a human verified them, and screening results carry the engine's written rationale.
- The same rate limit as the REST API applies: 120 requests per minute per key.