Haft MCP Server: Project Memory and Decision Records for Coding Agents

• Saves a non-binding note

  Stores a fact, observation, caveat or rationale in project memory, tied to the thing it is about and optionally given a freshness date.

  Try: “Note that the billing service rate-limits at 100 requests a minute.”

  Write · haft_note

• Frames a problem

  Writes down the problem being solved and the dimensions options will be compared on.

  Try: “Frame the problem of slow CI builds with cost and speed as dimensions.”

  Write · haft_problem

• Explores and compares options

  Records distinct solution variants with a diversity check, then compares them under the same ground rules.

  Try: “Compare Postgres, SQLite and DynamoDB for this service.”

  Write · haft_solution

• Records a decision

  Writes a decision with its invariants, claims, evidence and baseline. The README says it needs a direct, clear request from you.

  Try: “Record that we chose Postgres, with the benchmark as evidence.”

  Write · haft_decision

• Manages artifact lifecycle

  Scans, reviews, waives, reopens, supersedes or deprecates existing records so stale ones are flagged instead of silently trusted.

  Try: “Which of our decisions have gone stale?”

  Write · haft_refresh

• Searches project memory

  Searches and shows project status, the code graph and the stored records, and looks up the First Principles Framework source text.

  Try: “What did we decide about caching?”

  Read · haft_query

• Runs task checklists

  Opens a task-level method card with gates to pass before non-trivial work, then closes it with evidence or a recorded waiver.

  Try: “Start a method run for this refactor.”

  Write · haft_method

• Tracks delegated work

  Manages work commissions, the records that say what an outside runner may do and how that work moves through its lifecycle.

  Try: “Create a commission for the migration task.”

  Write · haft_commission

• Manages spec sections

  Handles typed spec sections stored in the project database. Approving, rebaselining and reopening a baseline are manual command-line steps.

  Try: “Show the spec section status.”

  Write · haft_spec_section

• Checks project setup

  Reads setup status and prepares a non-binding initial project profile for your review.

  Try: “What is the Haft setup status for this repo?”

  Write · haft_onboard

• Registers a thing of concern

  Creates one minimal, non-binding record for a named thing your work depends on, when the work gives a concrete reason.

  Try: “Register the payments service as an entity of concern.”

  Write · haft_entity

• Validates raw memory changes

  An expert tool that checks, and can admit, a raw batch of memory changes. Validation never admits changes by itself, and ordinary agents use haft_entity.

  Try: “Validate this memory change set.”

  Write · haft_memory

Watch out for

• It writes files in your project and home folder

  haft init changes your agent config, adds skills, and edits a managed section of CLAUDE.md or AGENTS.md. Only the content between the haft markers is touched. Review the diff before committing. github.com

• It can leave a large helper process running

  An optional embedding helper may start a local process and use about 1 to 2 GB of RAM. Set embedding.provider to none in ~/.haft/config.yaml to keep it off. github.com

• The installer is a script piped to bash

  The README installs with curl piped to bash. Read install.sh first if you prefer. github.com

Security advisories